queue.json

{
  "bridge": [
    {
      "arxiv_id": "2601.22705",
      "title": "CONCUR: High-Throughput Agentic Batch Inference of LLM via Congestion-Based Concurrency Control",
      "abstract": "Batch inference for agentic workloads stresses the GPU key-value (KV) cache in a sustained and cumulative manner, often causing severe throughput degradation well before memory capacity is exhausted. We identify this phenomenon as middle-phase thrashing, a previously under-characterized pathology in which cache efficiency collapses as long-lived agents accumulate state over time.   We argue that mitigating this pathology requires moving beyond reactive, request-level cache management to proactive, agent-level admission control. Drawing inspiration from congestion control in distributed systems, we view the KV cache as a shared resource whose efficient utilization depends on feedback-driven regulation. Based on this insight, we present CONCUR, a lightweight control layer that regulates agent admission to bound aggregate cache pressure while preserving execution continuity. CONCUR adapts a cache-aware control algorithm to dynamically adjust the number of active agents using runtime cache signals.   Across large models and real-world agent workloads, CONCUR prevents middle-phase thrashing and improves batch inference throughput by up to 4.09x on Qwen3-32B and 1.9x on DeepSeek-V3, while remaining compatible with existing LLM serving systems.",
      "authors": [
        "Qiaoling Chen",
        "Zhisheng Ye",
        "Tian Tang",
        "Peng Sun",
        "Boyu Tian",
        "Guoteng Wang",
        "Shenggui Li",
        "Yonggang Wen",
        "Zhenhua Han",
        "Tianwei Zhang"
      ],
      "published_at": "2026-01-30T08:27:20",
      "categories": [
        "cs.DC"
      ],
      "url": "http://arxiv.org/abs/2601.22705v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "systems",
      "domain_bucket": "llm",
      "paper_type": "systems",
      "narrow_domain_flag": false,
      "sim_public": 0.5403847694396973,
      "sim_memory": 0.5972424745559692,
      "sim_negative": 0.12155820429325104,
      "broad_relevance": 0.5403847694396973,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.40668225288391113,
      "evidence_score": 0.53,
      "direct_memory_relevance": 0.5972424745559692,
      "systems_leverage": 0.7,
      "deployment_proximity": 0.7,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.2,
      "transferability_score": 0.72,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.5731177687644959,
      "memory_score": 0.7281727423667907,
      "quality_score": 0.525,
      "bridge_score": 0.5731177687644959,
      "max_axis_score": 0.7281727423667907,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Memory/Storage Core",
        "Bridge",
        "Systems",
        "Inference"
      ],
      "status": "Cover now",
      "why_now": "Agentic inference is stressing serving stacks in ways standard request-level schedulers do not handle well, and this paper names a concrete new failure mode around sustained KV-cache pressure. The congestion-control framing gives operators a practical systems idea that could matter quickly as long-context and multi-step agent workloads scale.",
      "why_not_higher": "The contribution is focused on inference control-plane behavior rather than a broader shift in model architecture or training, so its audience is strongest among serving and systems practitioners. The abstract shows strong throughput wins, but it leaves open how robust the method is across diverse schedulers, latency targets, and production constraints.",
      "downgrade_reasons": [
        "agentic batch inference is still a narrower slice than general LLM serving",
        "throughput gains are clearer than latency, fairness, or deployment tradeoffs",
        "evidence in the abstract is limited to a small set of models and workloads"
      ],
      "what_would_raise_priority": "Independent results across more serving systems, hardware setups, and latency-sensitive workloads would make this a stronger must-cover paper.",
      "one_sentence_episode_hook": "What if LLM serving needs something like TCP congestion control, not just better kernels, to stop KV cache thrashing before GPUs look full?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2512.00719",
      "title": "SIMPLE: Disaggregating Sampling from GPU Inference into a Decision Plane for Faster Distributed LLM Serving",
      "abstract": "As large language models (LLMs) scale out with tensor parallelism (TP) and pipeline parallelism (PP) and production stacks have aggressively optimized the data plane (attention/GEMM and KV cache), sampling, the decision plane that turns logits into tokens, becomes a new bottleneck. This creates a structural holdout: sampling neither expands with TP nor balances across PP stages, so its share of iteration time grows as GPUs get faster and it caps pipeline frequency at the last stage. We present SIMPLE, a stage-agnostic, sequence-parallel, overlappable decision plane that disaggregates sampling into a CPU-side service and shrinks its runtime footprint back to a minor, hidden role. SIMPLE combines: (1) sequence-parallel sampling, which shards work along the batch dimension and removes vocabulary-axis collectives; (2) a CPU-based algorithm with column-wise penalties and truncation-first filtering to realize single-pass, linear-time kernels; and (3) speculative hot-vocab sampling (SHVS), which samples on a small hot set with rejection-correctness and uses a simple sizing model to choose the hot-vocab size that maximizes throughput. In evaluation, SIMPLE improves end-to-end throughput by up to 96% and reduces P95 latency by 20-65%. Crucially, SIMPLE requires no user-side code changes and composes with existing data-plane optimizations, unlocking scaling benefits that compound with future GPU generations.",
      "authors": [
        "Bohan Zhao",
        "Zane Cao",
        "Yongchao He"
      ],
      "published_at": "2025-11-30T04:15:34",
      "categories": [
        "cs.DC"
      ],
      "url": "http://arxiv.org/abs/2512.00719v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "systems",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.5261570811271667,
      "sim_memory": 0.5420535802841187,
      "sim_negative": 0.2318219542503357,
      "broad_relevance": 0.5261570811271667,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.5010901987552643,
      "evidence_score": 0.5,
      "direct_memory_relevance": 0.5420535802841187,
      "systems_leverage": 0.85,
      "deployment_proximity": 0.7,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.72,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.5530106541514397,
      "memory_score": 0.5386160740852356,
      "quality_score": 0.525,
      "bridge_score": 0.5386160740852356,
      "max_axis_score": 0.5530106541514397,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Bridge",
        "Systems",
        "Inference",
        "Memory/Storage Adjacent"
      ],
      "status": "Cover now",
      "why_now": "Distributed LLM serving has squeezed major gains out of attention, GEMM, and KV-cache plumbing, so control-path bottlenecks like sampling are becoming the next real limiter. A paper that reassigns that work off GPU and shows large end-to-end throughput and tail-latency gains is timely for anyone operating inference stacks.",
      "why_not_higher": "The memory/storage angle is real but secondary: this is mainly a serving-systems and decision-plane paper, not a direct KV-cache, paging, or offload mechanism paper. Its broad appeal also depends on whether the evaluation spans enough real models and deployment settings beyond the abstract's headline gains.",
      "downgrade_reasons": [
        "Memory connection is adjacent rather than core",
        "Abstract does not yet prove breadth across many serving stacks",
        "Could land as an optimization paper if implementation details dominate"
      ],
      "what_would_raise_priority": "Clear evidence that SIMPLE works across multiple production-grade LLM serving frameworks, model sizes, and heterogeneous hardware would raise it further.",
      "one_sentence_episode_hook": "What if the next big LLM serving speedup comes not from faster attention or smaller KV caches, but from kicking token sampling off the GPU entirely?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2602.00328",
      "title": "Harvest: Opportunistic Peer-to-Peer GPU Caching for LLM Inference",
      "abstract": "Large Language Model (LLM) inference is increasingly constrained by GPU memory capacity rather than compute throughput, driven by growing model sizes and the linear growth of the key-value (KV) cache during autoregressive decoding. Existing approaches mitigate memory pressure by offloading model state and KV tensors to host memory, but incur substantial latency due to limited PCIe bandwidth. We present Harvest, an opportunistic GPU cache management framework that exploits high-bandwidth peer-to-peer GPU interconnects to dynamically place model weights and KV cache in unused GPU memory. Harvest treats peer GPU memory as a transient cache tier, preserving correctness while reducing data movement overhead under dynamic memory availability. We demonstrate significant throughput speedup of more than 2 times by using Harvest to accelerate the retrieval of two widely-used inference components: expert layer weights and KV cache entries.",
      "authors": [
        "Nikhil Gopal",
        "Kostis Kaffes"
      ],
      "published_at": "2026-01-30T21:29:04",
      "categories": [
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2602.00328v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.437610387802124,
      "sim_memory": 0.5737738609313965,
      "sim_negative": 0.20268841087818146,
      "broad_relevance": 0.437610387802124,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.5027258396148682,
      "evidence_score": 0.45,
      "direct_memory_relevance": 0.5737738609313965,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 1.0,
      "transferability_score": 0.8799999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.5366919922828675,
      "memory_score": 0.685132158279419,
      "quality_score": 0.575,
      "bridge_score": 0.5366919922828675,
      "max_axis_score": 0.685132158279419,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Bridge",
        "Memory/Storage Core",
        "Systems",
        "Hardware",
        "Inference"
      ],
      "status": "Cover now",
      "why_now": "LLM serving is hitting memory-capacity and KV-cache limits right now, and this paper targets that bottleneck directly with a concrete systems mechanism. Using idle peer GPU memory as a transient cache tier is timely because multi-GPU inference stacks are already wrestling with offload bandwidth and KV movement costs.",
      "why_not_higher": "The win appears tied to specific hardware conditions: spare peer GPU memory and fast P2P interconnects, which narrows immediate applicability. From the abstract alone, the evaluation breadth across models, workloads, and production-like serving setups is not yet fully clear.",
      "downgrade_reasons": [
        "depends on unused peer GPU memory being available",
        "benefits may be topology- and interconnect-sensitive",
        "abstract does not show full production-serving coverage"
      ],
      "what_would_raise_priority": "Broader evidence across realistic serving workloads, heterogeneous cluster conditions, and multiple interconnect topologies would raise confidence and priority.",
      "one_sentence_episode_hook": "What if the fastest fix for KV-cache pressure is not CPU offload at all, but turning neighboring GPUs' idle memory into a live opportunistic cache?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2512.24449",
      "title": "PackKV: Reducing KV Cache Memory Footprint through LLM-Aware Lossy Compression",
      "abstract": "Transformer-based large language models (LLMs) have demonstrated remarkable potential across a wide range of practical applications. However, long-context inference remains a significant challenge due to the substantial memory requirements of the key-value (KV) cache, which can scale to several gigabytes as sequence length and batch size increase. In this paper, we present \\textbf{PackKV}, a generic and efficient KV cache management framework optimized for long-context generation. %, which synergistically supports both latency-critical and throughput-critical inference scenarios. PackKV introduces novel lossy compression techniques specifically tailored to the characteristics of KV cache data, featuring a careful co-design of compression algorithms and system architecture. Our approach is compatible with the dynamically growing nature of the KV cache while preserving high computational efficiency. Experimental results show that, under the same and minimum accuracy drop as state-of-the-art quantization methods, PackKV achieves, on average, \\textbf{153.2}\\% higher memory reduction rate for the K cache and \\textbf{179.6}\\% for the V cache. Furthermore, PackKV delivers extremely high execution throughput, effectively eliminating decompression overhead and accelerating the matrix-vector multiplication operation. Specifically, PackKV achieves an average throughput improvement of \\textbf{75.7}\\% for K and \\textbf{171.7}\\% for V across A100 and RTX Pro 6000 GPUs, compared to cuBLAS matrix-vector multiplication kernels, while demanding less GPU memory bandwidth. Code available on https://github.com/BoJiang03/PackKV",
      "authors": [
        "Bo Jiang",
        "Taolue Yang",
        "Youyuan Liu",
        "Xubin He",
        "Sheng Di",
        "Sian Jin"
      ],
      "published_at": "2025-12-30T20:05:32",
      "categories": [
        "cs.DC",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2512.24449v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "systems",
      "domain_bucket": "llm",
      "paper_type": "application",
      "narrow_domain_flag": false,
      "sim_public": 0.4577004909515381,
      "sim_memory": 0.7419873476028442,
      "sim_negative": 0.24648335576057434,
      "broad_relevance": 0.4577004909515381,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.3073432445526123,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.7419873476028442,
      "systems_leverage": 0.5499999999999999,
      "deployment_proximity": 0.7,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.4,
      "transferability_score": 0.9199999999999999,
      "clarity": 0.7,
      "reproducibility": 0.3,
      "public_interest_score": 0.4684116339683533,
      "memory_score": 0.7625962042808532,
      "quality_score": 0.505,
      "bridge_score": 0.4684116339683533,
      "max_axis_score": 0.7625962042808532,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Memory/Storage Core",
        "Systems",
        "Hardware",
        "Inference"
      ],
      "status": "Cover now",
      "why_now": "Long-context serving is still bottlenecked by KV cache footprint and bandwidth, so a method that compresses KV while also improving throughput is directly relevant to current deployment pain. This lands in an active wave of inference-systems work where concrete memory savings on modern GPUs matter immediately.",
      "why_not_higher": "The topic is highly relevant, but KV-cache compression is already a crowded lane, so novelty needs to clear a high bar. The abstract emphasizes benchmark gains over prior quantization baselines, but it is still unclear how broadly the results hold across models, sequence regimes, and production-serving stacks.",
      "downgrade_reasons": [
        "Crowded KV-cache optimization area",
        "Evidence in the abstract is benchmark-heavy",
        "Broad-audience appeal is lower than memory-lens appeal"
      ],
      "what_would_raise_priority": "Independent evidence across multiple frontier LLMs and real serving workloads would raise it further.",
      "one_sentence_episode_hook": "If KV cache is what makes long-context inference so expensive, PackKV asks a sharp systems question: can you throw away just enough cache information to save memory and bandwidth without throwing away model quality?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2603.01175",
      "title": "HAVEN: High-Bandwidth Flash Augmented Vector Engine for Large-Scale Approximate Nearest-Neighbor Search Acceleration",
      "abstract": "Retrieval-Augmented Generation (RAG) relies on large-scale Approximate Nearest Neighbor Search (ANNS) to retrieve semantically relevant context for large language models. Among ANNS methods, IVF-PQ offers an attractive balance between memory efficiency and search accuracy. However, achieving high recall requires reranking which fetches full-precision vectors for reranking, and the billion-scale vector databases need to reside in CPU DRAM or SSD due to the limited capacity of GPU HBM. This off-GPU data movement introduces substantial latency and throughput degradation.   We propose HAVEN, a GPU architecture augmented with High-Bandwidth Flash (HBF) which is a recently introduced die-stacked 3D NAND technology engineered to deliver terabyte-scale capacity and hundreds of GB/s read bandwidth. By integrating HBF and near-storage search unit as an on-package complement to HBM, HAVEN enables the full-precision vector database to reside entirely on-device, eliminating PCIe and DDR bottlenecks during reranking.   Through detailed modeling of re-architected 3D NAND subarrays, power-constrained HBF bandwidth, and end-to-end IVF-PQ pipelines, we demonstrate that HAVEN improves reranking throughput by up to 20x and latency up to 40x across billion-scale datasets compared to GPU-DRAM and GPU-SSD systems. Our results show that HBF-augmented GPUs enable high-recall retrieval at throughput previously achievable only without reranking, offering a promising direction for memory-centric AI accelerators.",
      "authors": [
        "Po-Kai Hsu",
        "Weihong Xu",
        "Qunyou Liu",
        "Tajana Rosing",
        "Shimeng Yu"
      ],
      "published_at": "2026-03-01T16:34:18",
      "categories": [
        "cs.AR",
        "cs.ET"
      ],
      "url": "http://arxiv.org/abs/2603.01175v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "hardware",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.32415521144866943,
      "sim_memory": 0.5237234830856323,
      "sim_negative": 0.17485398054122925,
      "broad_relevance": 0.32415521144866943,
      "momentum": 0.0,
      "teachability": 0.7,
      "novelty_score": 0.6018909811973572,
      "evidence_score": 0.34,
      "direct_memory_relevance": 0.5237234830856323,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.5,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 1.0,
      "transferability_score": 0.64,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.4475302106142044,
      "memory_score": 0.7421170449256896,
      "quality_score": 0.545,
      "bridge_score": 0.4475302106142044,
      "max_axis_score": 0.7421170449256896,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Memory/Storage Core",
        "Bridge",
        "Systems",
        "Hardware",
        "Inference"
      ],
      "status": "Cover now",
      "why_now": "RAG and vector retrieval remain live serving bottlenecks, and this paper attacks the exact bandwidth and capacity wall that appears when reranking billion-scale indexes. It is a direct memory-centric architecture proposal with clear relevance to how future retrieval systems could be built for LLM inference stacks.",
      "why_not_higher": "The paper is still a specialized ANN accelerator story rather than a change to mainstream model training or serving across the whole AI stack. The evidence is based on detailed architectural modeling rather than deployed hardware, which limits immediate confidence.",
      "downgrade_reasons": [
        "Narrowly centered on IVF-PQ reranking for ANN search",
        "Modeled architecture rather than shipping system evidence",
        "Broad-audience appeal is lower than for general LLM serving papers"
      ],
      "what_would_raise_priority": "A stronger cross-system comparison showing real end-to-end gains on production RAG workloads or broader applicability beyond IVF-PQ reranking would raise it.",
      "one_sentence_episode_hook": "What if the real bottleneck in RAG is not the model but the trip full-precision vectors take through the memory hierarchy, and fixing it requires putting flash on the GPU package itself?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.22001",
      "title": "Heterogeneous Computing: The Key to Powering the Future of AI Agent Inference",
      "abstract": "AI agent inference is driving an inference heavy datacenter future and exposes bottlenecks beyond compute - especially memory capacity, memory bandwidth and high-speed interconnect. We introduce two metrics - Operational Intensity (OI) and Capacity Footprint (CF) - that jointly explain regimes the classic roofline analysis misses, including the memory capacity wall. Across agentic workflows (chat, coding, web use, computer use) and base model choices (GQA/MLA, MoE, quantization), OI/CF can shift dramatically, with long context KV cache making decode highly memory bound. These observations motivate disaggregated serving and system level heterogeneity: specialized prefill and decode accelerators, broader scale up networking, and decoupled compute-memory enabled by optical I/O. We further hypothesize agent-hardware co design, multiple inference accelerators within one system, and high bandwidth, large capacity memory disaggregation as foundations for adaptation to evolving OI/CF. Together, these directions chart a path to sustain efficiency and capability for large scale agentic AI inference.",
      "authors": [
        "Yiren Zhao",
        "Junyi Liu"
      ],
      "published_at": "2026-01-29T17:11:46",
      "categories": [
        "cs.AI",
        "cs.AR",
        "cs.DC"
      ],
      "url": "http://arxiv.org/abs/2601.22001v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.4255807399749756,
      "sim_memory": 0.6131550073623657,
      "sim_negative": 0.19539843499660492,
      "broad_relevance": 0.4255807399749756,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.45346856117248535,
      "evidence_score": 0.32,
      "direct_memory_relevance": 0.6131550073623657,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.25,
      "bandwidth_capacity": 0.8,
      "transferability_score": 0.7,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.5256945061683655,
      "memory_score": 0.7189465022087096,
      "quality_score": 0.525,
      "bridge_score": 0.5256945061683655,
      "max_axis_score": 0.7189465022087096,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Memory/Storage Core",
        "Bridge",
        "Systems",
        "Hardware",
        "Inference"
      ],
      "status": "Monitor",
      "why_now": "Agentic inference is exposing memory-capacity, bandwidth, and interconnect limits that standard compute-centric framing misses, so a paper that centers those bottlenecks is timely. Its OI/CF lens also matches the industry's shift toward long-context serving, disaggregation, and heterogeneous inference stacks.",
      "why_not_higher": "From the abstract, this looks partly like a systems framing and agenda-setting paper rather than a fully validated deployment study with hard comparative evidence. Several key claims are still stated as hypotheses about future architectures rather than demonstrated wins on production-grade end-to-end systems.",
      "downgrade_reasons": [
        "Evidence appears stronger on diagnosis than on validated system improvements",
        "Some architecture recommendations seem forward-looking or speculative",
        "May overlap with existing discussion around disaggregated serving and KV-cache-driven memory bottlenecks"
      ],
      "what_would_raise_priority": "A stronger priority case would come from clear measurements on realistic agent workloads showing OI/CF predicts system choices better than existing serving heuristics and leads to demonstrated efficiency gains.",
      "one_sentence_episode_hook": "If AI agents are turning inference into a memory-and-interconnect problem, this paper argues the next breakthrough is not one better chip but heterogeneous systems built around the capacity wall.",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2510.27257",
      "title": "Synergistic Tensor and Pipeline Parallelism",
      "abstract": "In the machine learning system, the hybrid model parallelism combining tensor parallelism (TP) and pipeline parallelism (PP) has become the dominant solution for distributed training of Large Language Models~(LLMs) and Multimodal LLMs (MLLMs). However, TP introduces significant collective communication overheads, while PP suffers from synchronization inefficiencies such as pipeline bubbles. Existing works primarily address these challenges from isolated perspectives, focusing either on overlapping TP communication or on flexible PP scheduling to mitigate pipeline bubbles. In this paper, we propose a new synergistic tensor and pipeline parallelism schedule that simultaneously reduces both types of bubbles. Our proposed schedule decouples the forward and backward passes in PP into fine-grained computation units, which are then braided to form a composite computation sequence. This compositional structure enables near-complete elimination of TP-related bubbles. Building upon this structure, we further design the PP schedule to minimize PP bubbles. Experimental results demonstrate that our approach improves training throughput by up to 12% for LLMs and 16% for MLLMs compared to existing scheduling methods. Our source code is avaiable at https://github.com/MICLAB-BUPT/STP.",
      "authors": [
        "Mengshi Qi",
        "Jiaxuan Peng",
        "Jie Zhang",
        "Juan Zhu",
        "Yong Li",
        "Huadong Ma"
      ],
      "published_at": "2025-10-31T07:53:40",
      "categories": [
        "cs.DC"
      ],
      "url": "http://arxiv.org/abs/2510.27257v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "systems",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.6211696863174438,
      "sim_memory": 0.6670984625816345,
      "sim_negative": 0.28806158900260925,
      "broad_relevance": 0.6211696863174438,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.5031062960624695,
      "evidence_score": 0.48000000000000004,
      "direct_memory_relevance": 0.6670984625816345,
      "systems_leverage": 0.7,
      "deployment_proximity": 0.7,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6699999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5518168503046036,
      "memory_score": 0.5191295387744903,
      "quality_score": 0.545,
      "bridge_score": 0.5191295387744903,
      "max_axis_score": 0.5518168503046036,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Bridge",
        "Systems",
        "Training",
        "Memory/Storage Adjacent"
      ],
      "status": "Monitor",
      "why_now": "Large-model training is still constrained by communication and synchronization overheads, so a schedule that jointly attacks tensor-parallel and pipeline bubbles is relevant to current frontier training stacks. The paper is timely because hybrid TP+PP remains standard for scaling both LLMs and multimodal models.",
      "why_not_higher": "This is a training-systems optimization paper, not a broadly legible shift for most AI practitioners, and the abstract emphasizes throughput gains without enough hardware and workload detail to judge how robust they are across stacks. Its memory angle is real but indirect: the core contribution is parallel scheduling and communication overlap rather than cache, offload, paging, or storage mechanisms.",
      "downgrade_reasons": [
        "Broad appeal is limited outside distributed training practitioners",
        "Memory/storage connection is adjacent rather than core",
        "Abstract lacks enough detail on hardware realism and generality of the reported gains",
        "Risks reading as an incremental scheduler improvement rather than a major conceptual shift"
      ],
      "what_would_raise_priority": "Clear evidence that the schedule transfers across widely used training frameworks, model scales, and interconnect regimes with consistent end-to-end gains would raise it.",
      "one_sentence_episode_hook": "What if the next training speedup for giant models comes not from new hardware, but from braiding tensor and pipeline parallelism so the bubbles in each cancel the other?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.11577",
      "title": "Computation-Bandwidth-Memory Trade-offs: A Unified Paradigm for AI Infrastructure",
      "abstract": "Large-scale artificial intelligence models are transforming industries and redefining human machine collaboration. However, continued scaling exposes critical limitations in hardware, including constraints on computation, bandwidth, and memory. These dimensions are tightly interconnected, so improvements in one often create bottlenecks in others, making isolated optimizations less effective. Balancing them to maximize system efficiency remains a central challenge in scalable AI design. To address this challenge, we introduce {Computation-Bandwidth-Memory Trade-offs}, termed the {AI Trinity}, a unified paradigm that positions {computation}, {bandwidth}, and {memory} as coequal pillars for next-generation AI infrastructure. AI Trinity enables dynamic allocation of resources across these pillars, alleviating single-resource bottlenecks and adapting to diverse scenarios to optimize system performance. Within this framework, AI Trinity identifies three fundamental trade-offs: (1) {More Computation$\\rightarrow$Less Bandwidth}, wherein computational resources are exploited to reduce data transmission under limited bandwidth conditions, (2) {More Bandwidth$\\rightarrow$Less Memory}, which exploits abundant communication capacity to populate or refresh memory when local storage resources are constrained, and (3) {More Memory$\\rightarrow$Less Computation}, whereby storage capacity are utilized to mitigate redundant computation when computational costs are prohibitive. We illustrate the effectiveness of AI Trinity through representative system designs spanning edge-cloud communication, large-scale distributed training, and model inference. The innovations embodied in AI Trinity advance a new paradigm for scalable AI infrastructure, providing both a conceptual foundation and practical guidance for a broad range of application scenarios.",
      "authors": [
        "Yuankai Fan",
        "Qizhen Weng",
        "Xuelong Li"
      ],
      "published_at": "2025-12-30T17:35:14",
      "categories": [
        "cs.DC"
      ],
      "url": "http://arxiv.org/abs/2601.11577v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "systems",
      "domain_bucket": "robotics",
      "paper_type": "systems",
      "narrow_domain_flag": false,
      "sim_public": 0.4794940948486328,
      "sim_memory": 0.5329809188842773,
      "sim_negative": 0.17475129663944244,
      "broad_relevance": 0.4794940948486328,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.5189214944839478,
      "evidence_score": 0.28,
      "direct_memory_relevance": 0.5329809188842773,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.7,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.4,
      "transferability_score": 0.6799999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.491686452627182,
      "memory_score": 0.5848942756652832,
      "quality_score": 0.525,
      "bridge_score": 0.491686452627182,
      "max_axis_score": 0.5848942756652832,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Memory/Storage Core",
        "Bridge",
        "Systems",
        "Hardware",
        "Training",
        "Inference"
      ],
      "status": "Monitor",
      "why_now": "As frontier models keep running into serving and training bottlenecks, a paper that frames compute, bandwidth, and memory as a single design space is timely for both infrastructure builders and a broad AI audience. It also maps well onto current pressure around KV cache growth, interconnect limits, and system-level efficiency.",
      "why_not_higher": "This reads more like a unifying perspective paper than a decisive new systems result with strong empirical proof on realistic end-to-end deployments. The abstract promises representative examples, but not enough concrete workload, hardware, or benchmark evidence to rank it as an immediate must-cover episode.",
      "downgrade_reasons": [
        "Conceptual framing outweighs demonstrated system gains",
        "Unclear empirical depth on realistic workloads and hardware",
        "May overlap with existing AI systems bottleneck discussions without a sharp new mechanism"
      ],
      "what_would_raise_priority": "Clear quantitative results on modern LLM training or inference stacks showing when each trade-off wins in practice would raise it substantially.",
      "one_sentence_episode_hook": "What if the real bottleneck in AI is not compute, memory, or bandwidth alone, but our failure to treat them as one coupled system?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2603.28239",
      "title": "A Switch-Centric In-Network Architecture for Accelerating LLM Inference in Shared-Memory Network",
      "abstract": "In-network computing techniques, exemplified by NVLink Sharp (NVLS), offer a promising approach to addressing the communication bottlenecks in LLM inference by offloading collective operations, such as All-Reduce, to switches. However, the accelerator-centric architecture of NVLS suffers from two fundamental limitations: 1) it relies on GPU load instructions to trigger reduction operations, which means that the data reduced in the switch must be additionally transferred back to the initiating GPU rather than being broadcast directly, thereby introducing unnecessary communication overhead; 2) due to its architectural constraints, NVLS cannot offload operators that are not decomposable into memory-semantic instructions, such as the in-network quantization (INQ) proposed in this work. As a result, All-Reduce in NVLS must operate at FP16/BF16 precision, leading to substantial bandwidth waste.To address these limitations, we propose SCIN, the first switch-centric in-network architecture for shared-memory networks of AI accelerators, enabling both low-latency and high-bandwidth All-Reduce. Specifically, we introduce an in-switch accelerator (ISA) capable of initiating memory-semantic operations for in-network processing, together with a co-designed communication fabric that incurs negligible protocol overhead. By eliminating redundant data movement, SCIN delivers lower All-Reduce latency than NVLS. Moreover, by integrating a quantization module into the ISA, SCIN enables INQ for All-Reduce, reducing its precision to 8 bits and nearly doubling bandwidth with negligible accuracy loss. We also present a prototype of SCIN on a multi-FPGA system to demonstrate its feasibility and effectiveness. Experimental results show that our design accelerates All-Reduce by up to 8.7x for small messages and 3.8x for large messages, leading up to 1.74x faster TTFT and 1.34x faster TPOT on LLaMA-2 models.",
      "authors": [
        "Aojie Jiang",
        "Kang Zhu",
        "Zhiheng Zhang",
        "Zhengxu Su",
        "Juntao Liu",
        "Yuan Du",
        "Li Du"
      ],
      "published_at": "2026-03-30T09:59:11",
      "categories": [
        "cs.AR"
      ],
      "url": "http://arxiv.org/abs/2603.28239v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "hardware",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.46526843309402466,
      "sim_memory": 0.521072506904602,
      "sim_negative": 0.18257898092269897,
      "broad_relevance": 0.46526843309402466,
      "momentum": 0.0,
      "teachability": 0.7,
      "novelty_score": 0.5278318524360657,
      "evidence_score": 0.48000000000000004,
      "direct_memory_relevance": 0.521072506904602,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.5,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.6000000000000001,
      "transferability_score": 0.63,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.4787553077936172,
      "memory_score": 0.6613217520713807,
      "quality_score": 0.545,
      "bridge_score": 0.4787553077936172,
      "max_axis_score": 0.6613217520713807,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "30d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Memory/Storage Core",
        "Bridge",
        "Systems",
        "Hardware",
        "Inference"
      ],
      "status": "Monitor",
      "why_now": "LLM serving is increasingly bottlenecked by collective communication, and this paper targets that pain directly with a switch-centric design plus in-network quantization. It matters now because inference latency and bandwidth efficiency are becoming first-order deployment constraints for larger shared-memory accelerator systems.",
      "why_not_higher": "The idea is important, but the paper is still heavily architecture- and prototype-driven rather than something most AI teams can act on soon. Its impact is strongest for specialized high-end serving stacks, not the broad median practitioner.",
      "downgrade_reasons": [
        "specialized hardware dependency",
        "prototype evidence rather than production deployment",
        "limited immediate applicability for most AI builders",
        "episode risk: could skew into benchmark-heavy architecture detail"
      ],
      "what_would_raise_priority": "A stronger priority case would come from evidence on production-like GPU clusters or adoption paths showing the design transfers beyond a custom multi-FPGA prototype.",
      "one_sentence_episode_hook": "What if the fastest way to serve LLMs is to stop treating the network switch as plumbing and turn it into an active participant in all-reduce and quantization?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2510.26730",
      "title": "ExpertFlow: Adaptive Expert Scheduling and Memory Coordination for Efficient MoE Inference",
      "abstract": "The expansion of large language models is increasingly limited by the constrained memory capacity of modern GPUs. To mitigate this, Mixture-of-Experts (MoE) architectures activate only a small portion of parameters during inference, significantly lowering both memory demand and computational overhead. However, conventional MoE inference approaches, which select active experts independently at each layer, often introduce considerable latency because of frequent parameter transfers between host and GPU memory. In addition, current cross-layer prediction strategies, which are typically based on fixed steps, lack adaptability across different hardware platforms and workloads, thereby reducing their robustness and effectiveness.   To address these challenges, we present ExpertFlow, a runtime system for MoE inference that combines adaptive expert prefetching and cache-aware routing. ExpertFlow continuously adjusts its prediction horizon for expert activation by leveraging runtime statistics such as transfer bandwidth, parameter dimensionality, and model feedback signals. Furthermore, it incorporates a hybrid cross-layer prediction scheme that fuses pregating information with intermediate computational states to anticipate future expert needs. By adaptively refining prefetching decisions and aligning them with actual usage behavior, ExpertFlow effectively decreases cache misses and removes latency caused by expert swap-ins. Our evaluation demonstrates that ExpertFlow reduces model stall time to less than 0.1% of the baseline, highlighting its capability to optimize MoE inference under stringent memory constraints.",
      "authors": [
        "Zixu Shen",
        "Kexin Chu",
        "Yifan Zhang",
        "Dawei Xiang",
        "Runxin Wu",
        "Wei Zhang"
      ],
      "published_at": "2025-10-30T17:29:27",
      "categories": [
        "cs.DC",
        "cs.AI",
        "cs.PF"
      ],
      "url": "http://arxiv.org/abs/2510.26730v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "systems",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.5477154850959778,
      "sim_memory": 0.5541058778762817,
      "sim_negative": 0.27488723397254944,
      "broad_relevance": 0.5477154850959778,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.4982467293739319,
      "evidence_score": 0.34,
      "direct_memory_relevance": 0.5541058778762817,
      "systems_leverage": 0.5499999999999999,
      "deployment_proximity": 0.7,
      "memory_adjacent_future_value": 0.5,
      "bandwidth_capacity": 0.4,
      "transferability_score": 0.69,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.4790516549348831,
      "memory_score": 0.7082317633628845,
      "quality_score": 0.525,
      "bridge_score": 0.4790516549348831,
      "max_axis_score": 0.7082317633628845,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Memory/Storage Core",
        "Systems",
        "Inference",
        "Bridge"
      ],
      "status": "Monitor",
      "why_now": "MoE deployment keeps running into GPU memory limits and expert swap latency, so adaptive prefetching and cache-aware scheduling are timely systems problems. This paper sits directly on the current serve-more-model-with-less-HBM pressure point.",
      "why_not_higher": "The paper looks operationally relevant but still fairly MoE-serving specific rather than broadly changing mainstream LLM practice. The abstract also does not yet show enough workload and hardware detail to tell whether the gains hold beyond a favorable setup.",
      "downgrade_reasons": [
        "MoE inference is important but still narrower than general LLM serving",
        "Evidence in the abstract is promising but underspecified",
        "Could end up as a strong optimization result rather than a field-shifting method"
      ],
      "what_would_raise_priority": "Clear results on widely used MoE models and heterogeneous real serving hardware, with strong latency-throughput tradeoff comparisons against current MoE runtime baselines, would raise it.",
      "one_sentence_episode_hook": "If MoE models are supposed to save memory, why do they still stall on expert swaps, and can adaptive prefetching finally make sparse models behave like fast ones in production?",
      "source": "digest",
      "added": "",
      "issue_number": null
    }
  ],
  "public": [
    {
      "arxiv_id": "2512.02010",
      "title": "Four Over Six: More Accurate NVFP4 Quantization with Adaptive Block Scaling",
      "abstract": "As large language models have grown larger, interest has grown in low-precision numerical formats such as NVFP4 as a way to improve speed and reduce memory usage. However, quantizing models to NVFP4 remains difficult as the lack of precision generally degrades model performance. In this work, we address this issue with Four Over Six (4/6), a modification to the block-scaled NVFP4 quantization algorithm that yields reduced quantization error. Unlike integer formats, floating point formats have non-uniform step sizes which create larger quantization error on larger values. 4/6 takes advantage of this by adaptively scaling some blocks to smaller FP4 values, making the distribution of representable values more uniform and reducing quantization error for near-maximal values. We show that 4/6 can be implemented efficiently on NVIDIA Blackwell GPUs, resulting in performance gains during both pre-training and inference with minimal computational overhead. In pre-training experiments with the Nemotron 3 Nano 30B-A3B model architecture, we find that 4/6 brings training loss closer to BF16 compared to models trained with current state-of-the-art NVFP4 training recipes. Our code is available at http://github.com/mit-han-lab/fouroversix.",
      "authors": [
        "Jack Cook",
        "Junxian Guo",
        "Guangxuan Xiao",
        "Yujun Lin",
        "Song Han"
      ],
      "published_at": "2025-12-01T18:59:45",
      "categories": [
        "cs.CL",
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2512.02010v3",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.5141356587409973,
      "sim_memory": 0.6465221047401428,
      "sim_negative": 0.23632439970970154,
      "broad_relevance": 0.5141356587409973,
      "momentum": 0.7500000000000001,
      "teachability": 0.85,
      "novelty_score": 0.5399790108203888,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.6465221047401428,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.6152375492453577,
      "memory_score": 0.2899566314220428,
      "quality_score": 0.575,
      "bridge_score": 0.2899566314220428,
      "max_axis_score": 0.6152375492453577,
      "influencer_boost": 0.225,
      "influencer_matches": [
        "Song Han"
      ],
      "pwc_trending_flag": false,
      "social_score": 0.225,
      "scoring_sources": [
        "influencer:Song Han",
        "social_momentum",
        "compound_window:180d"
      ],
      "time_window": "180d",
      "compound_window_boost": 0.045,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2603.26639",
      "title": "Make Geometry Matter for Spatial Reasoning",
      "abstract": "Empowered by large-scale training, vision-language models (VLMs) achieve strong image and video understanding, yet their ability to perform spatial reasoning in both static scenes and dynamic videos remains limited. Recent advances try to handle this limitation by injecting geometry tokens from pretrained 3D foundation models into VLMs. Nevertheless, we observe that naive token fusion followed by standard fine-tuning in this line of work often leaves such geometric cues underutilized for spatial reasoning, as VLMs tend to rely heavily on 2D visual cues. In this paper, we propose GeoSR, a framework designed to make geometry matter by encouraging VLMs to actively reason with geometry tokens. GeoSR introduces two key components: (1) Geometry-Unleashing Masking, which strategically masks portions of 2D vision tokens during training to weaken non-geometric shortcuts and force the model to consult geometry tokens for spatial reasoning; and (2) Geometry-Guided Fusion, a gated routing mechanism that adaptively amplifies geometry token contributions in regions where geometric evidence is critical. Together, these designs unleash the potential of geometry tokens for spatial reasoning tasks. Extensive experiments on both static and dynamic spatial reasoning benchmarks demonstrate that GeoSR consistently outperforms prior methods and establishes new state-of-the-art performance by effectively leveraging geometric information. The project page is available at https://suhzhang.github.io/GeoSR/.",
      "authors": [
        "Shihua Zhang",
        "Qiuhong Shen",
        "Shizun Wang",
        "Tianbo Pan",
        "Xinchao Wang"
      ],
      "published_at": "2026-03-27T17:45:12",
      "categories": [
        "cs.CV",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2603.26639v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": true,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "application",
      "domain_bucket": "vision",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.3662530779838562,
      "sim_memory": 0.2911671996116638,
      "sim_negative": 0.4095418453216553,
      "broad_relevance": 0.3662530779838562,
      "momentum": 1.0,
      "teachability": 0.7,
      "novelty_score": 0.5752736330032349,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.2911671996116638,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.25,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5961669683456421,
      "memory_score": 0.20835015988349911,
      "quality_score": 0.545,
      "bridge_score": 0.20835015988349911,
      "max_axis_score": 0.5961669683456421,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": true,
      "social_score": 0.1,
      "scoring_sources": [
        "pwc_trending",
        "social_momentum",
        "pwc_trending"
      ],
      "time_window": "30d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.00844",
      "title": "Value-guided action planning with JEPA world models",
      "abstract": "Building deep learning models that can reason about their environment requires capturing its underlying dynamics. Joint-Embedded Predictive Architectures (JEPA) provide a promising framework to model such dynamics by learning representations and predictors through a self-supervised prediction objective. However, their ability to support effective action planning remains limited. We propose an approach to enhance planning with JEPA world models by shaping their representation space so that the negative goal-conditioned value function for a reaching cost in a given environment is approximated by a distance (or quasi-distance) between state embeddings. We introduce a practical method to enforce this constraint during training and show that it leads to significantly improved planning performance compared to standard JEPA models on simple control tasks.",
      "authors": [
        "Matthieu Destrade",
        "Oumayma Bounou",
        "Quentin Le Lidec",
        "Jean Ponce",
        "Yann LeCun"
      ],
      "published_at": "2025-12-28T20:17:49",
      "categories": [
        "cs.LG",
        "cs.AI",
        "cs.RO"
      ],
      "url": "http://arxiv.org/abs/2601.00844v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "robotics",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.3781801462173462,
      "sim_memory": 0.3515470027923584,
      "sim_negative": 0.3398587107658386,
      "broad_relevance": 0.3781801462173462,
      "momentum": 0.7833333333333333,
      "teachability": 0.85,
      "novelty_score": 0.4997263550758362,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.3515470027923584,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.575079663793246,
      "memory_score": 0.2014641008377075,
      "quality_score": 0.545,
      "bridge_score": 0.2014641008377075,
      "max_axis_score": 0.575079663793246,
      "influencer_boost": 0.2375,
      "influencer_matches": [
        "Yann LeCun"
      ],
      "pwc_trending_flag": false,
      "social_score": 0.2375,
      "scoring_sources": [
        "influencer:Yann LeCun",
        "social_momentum",
        "compound_window:180d"
      ],
      "time_window": "180d",
      "compound_window_boost": 0.045,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2603.27027",
      "title": "TAPS: Task Aware Proposal Distributions for Speculative Sampling",
      "abstract": "Speculative decoding accelerates autoregressive generation by letting a lightweight draft model propose future tokens that a larger target model then verifies in parallel. In practice, however, draft models are usually trained on broad generic corpora, which leaves it unclear how much speculative decoding quality depends on the draft training distribution. We study this question with lightweight HASS and EAGLE-2 drafters trained on MathInstruct, ShareGPT, and mixed-data variants, evaluated on MT-Bench, GSM8K, MATH-500, and SVAMP. Measured by acceptance length, task-specific training yields clear specialization: MathInstruct-trained drafts are strongest on reasoning benchmarks, while ShareGPT-trained drafts are strongest on MT-Bench. Mixed-data training improves robustness, but larger mixtures do not dominate across decoding temperatures. We also study how to combine specialized drafters at inference time. Naive checkpoint averaging performs poorly, whereas confidence-based routing improves over single-domain drafts and merged-tree verification yields the highest acceptance length overall for both backbones. Finally, confidence is a more useful routing signal than entropy: rejected tokens tend to have higher entropy, but confidence produces much clearer benchmark-level routing decisions. These results show that speculative decoding quality depends not only on draft architecture, but also on the match between draft training data and downstream workload, and that specialized drafters are better combined at inference time than in weight space.",
      "authors": [
        "Mohamad Zbib",
        "Mohamad Bazzi",
        "Ammar Mohanna",
        "Hasan Abed Al Kader Hammoud",
        "Bernard Ghanem"
      ],
      "published_at": "2026-03-27T22:34:20",
      "categories": [
        "cs.CL",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2603.27027v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": true,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.694627046585083,
      "sim_memory": 0.5123721361160278,
      "sim_negative": 0.2292865514755249,
      "broad_relevance": 0.694627046585083,
      "momentum": 0.4,
      "teachability": 0.75,
      "novelty_score": 0.5307408571243286,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.5123721361160278,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.5,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.5779992425441742,
      "memory_score": 0.29971164083480833,
      "quality_score": 0.525,
      "bridge_score": 0.29971164083480833,
      "max_axis_score": 0.5779992425441742,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "30d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2512.24497",
      "title": "What Drives Success in Physical Planning with Joint-Embedding Predictive World Models?",
      "abstract": "A long-standing challenge in AI is to develop agents capable of solving a wide range of physical tasks and generalizing to new, unseen tasks and environments. A popular recent approach involves training a world model from state-action trajectories and subsequently use it with a planning algorithm to solve new tasks. Planning is commonly performed in the input space, but a recent family of methods has introduced planning algorithms that optimize in the learned representation space of the world model, with the promise that abstracting irrelevant details yields more efficient planning. In this work, we characterize models from this family as JEPA-WMs and investigate the technical choices that make algorithms from this class work. We propose a comprehensive study of several key components with the objective of finding the optimal approach within the family. We conducted experiments using both simulated environments and real-world robotic data, and studied how the model architecture, the training objective, and the planning algorithm affect planning success. We combine our findings to propose a model that outperforms two established baselines, DINO-WM and V-JEPA-2-AC, in both navigation and manipulation tasks. Code, data and checkpoints are available at https://github.com/facebookresearch/jepa-wms.",
      "authors": [
        "Basile Terver",
        "Tsung-Yen Yang",
        "Jean Ponce",
        "Adrien Bardes",
        "Yann LeCun"
      ],
      "published_at": "2025-12-30T22:50:03",
      "categories": [
        "cs.AI",
        "cs.LG",
        "cs.RO",
        "stat.ML"
      ],
      "url": "http://arxiv.org/abs/2512.24497v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "robotics",
      "paper_type": "application",
      "narrow_domain_flag": false,
      "sim_public": 0.4279174208641052,
      "sim_memory": 0.3847622871398926,
      "sim_negative": 0.3060411810874939,
      "broad_relevance": 0.4279174208641052,
      "momentum": 0.7833333333333333,
      "teachability": 0.85,
      "novelty_score": 0.5783583223819733,
      "evidence_score": 0.3,
      "direct_memory_relevance": 0.3847622871398926,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.3,
      "public_interest_score": 0.5867956412831943,
      "memory_score": 0.19942868614196776,
      "quality_score": 0.47500000000000003,
      "bridge_score": 0.19942868614196776,
      "max_axis_score": 0.5867956412831943,
      "influencer_boost": 0.2375,
      "influencer_matches": [
        "Yann LeCun"
      ],
      "pwc_trending_flag": false,
      "social_score": 0.2375,
      "scoring_sources": [
        "influencer:Yann LeCun",
        "social_momentum",
        "compound_window:180d"
      ],
      "time_window": "180d",
      "compound_window_boost": 0.045,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2602.23643",
      "title": "AI Must Embrace Specialization via Superhuman Adaptable Intelligence",
      "abstract": "Everyone from AI executives and researchers to doomsayers, politicians, and activists is talking about Artificial General Intelligence (AGI). Yet, they often don't seem to agree on its exact definition. One common definition of AGI is an AI that can do everything a human can do, but are humans truly general? In this paper, we address what's wrong with our conception of AGI, and why, even in its most coherent formulation, it is a flawed concept to describe the future of AI. We explore whether the most widely accepted definitions are plausible, useful, and truly general. We argue that AI must embrace specialization, rather than strive for generality, and in its specialization strive for superhuman performance, and introduce Superhuman Adaptable Intelligence (SAI). SAI is defined as intelligence that can learn to exceed humans at anything important that we can do, and that can fill in the skill gaps where humans are incapable. We then lay out how SAI can help hone a discussion around AI that was blurred by an overloaded definition of AGI, and extrapolate the implications of using it as a guide for the future.",
      "authors": [
        "Judah Goldfeder",
        "Philippe Wyder",
        "Yann LeCun",
        "Ravid Shwartz Ziv"
      ],
      "published_at": "2026-02-27T03:26:21",
      "categories": [
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2602.23643v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.4145880937576294,
      "sim_memory": 0.264610230922699,
      "sim_negative": 0.30329158902168274,
      "broad_relevance": 0.4145880937576294,
      "momentum": 0.7383333333333334,
      "teachability": 0.75,
      "novelty_score": 0.5178835988044739,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.264610230922699,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.5597256346146267,
      "memory_score": 0.1753830692768097,
      "quality_score": 0.575,
      "bridge_score": 0.1753830692768097,
      "max_axis_score": 0.5597256346146267,
      "influencer_boost": 0.2375,
      "influencer_matches": [
        "Yann LeCun"
      ],
      "pwc_trending_flag": false,
      "social_score": 0.2375,
      "scoring_sources": [
        "influencer:Yann LeCun",
        "social_momentum",
        "compound_window:90d"
      ],
      "time_window": "90d",
      "compound_window_boost": 0.0315,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2602.00359",
      "title": "Position: Agentic Evolution is the Path to Evolving LLMs",
      "abstract": "As Large Language Models (LLMs) move from curated training sets into open-ended real-world environments, a fundamental limitation emerges: static training cannot keep pace with continual deployment environment change. Scaling training-time and inference-time compute improves static capability but does not close this train-deploy gap. We argue that addressing this limitation requires a new scaling axis-evolution. Existing deployment-time adaptation methods, whether parametric fine-tuning or heuristic memory accumulation, lack the strategic agency needed to diagnose failures and produce durable improvements. Our position is that agentic evolution represents the inevitable future of LLM adaptation, elevating evolution itself from a fixed pipeline to an autonomous evolver agent. We instantiate this vision in a general framework, A-Evolve, which treats deployment-time improvement as a deliberate, goal-directed optimization process over persistent system state. We further propose the evolution-scaling hypothesis: the capacity for adaptation scales with the compute allocated to evolution, positioning agentic evolution as a scalable path toward sustained, open-ended adaptation in the real world. Our code is publicly available at https://github.com/A-EVO-Lab/a-evolve.",
      "authors": [
        "Minhua Lin",
        "Hanqing Lu",
        "Zhan Shi",
        "Bing He",
        "Rui Mao",
        "Zhiwei Zhang",
        "Zongyu Wu",
        "Xianfeng Tang",
        "Hui Liu",
        "Zhenwei Dai"
      ],
      "published_at": "2026-01-30T22:15:58",
      "categories": [
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2602.00359v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "application",
      "narrow_domain_flag": false,
      "sim_public": 0.4625900983810425,
      "sim_memory": 0.41107407212257385,
      "sim_negative": 0.254189133644104,
      "broad_relevance": 0.4625900983810425,
      "momentum": 0.81,
      "teachability": 0.75,
      "novelty_score": 0.4633867144584656,
      "evidence_score": 0.3,
      "direct_memory_relevance": 0.41107407212257385,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7,
      "reproducibility": 0.3,
      "public_interest_score": 0.5652850366830826,
      "memory_score": 0.26132222163677216,
      "quality_score": 0.505,
      "bridge_score": 0.26132222163677216,
      "max_axis_score": 0.5652850366830826,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": true,
      "social_score": 0.1,
      "scoring_sources": [
        "pwc_trending",
        "social_momentum",
        "pwc_trending",
        "compound_window:90d"
      ],
      "time_window": "90d",
      "compound_window_boost": 0.063,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2602.09040",
      "title": "Soft Clustering Anchors for Self-Supervised Speech Representation Learning in Joint Embedding Prediction Architectures",
      "abstract": "Joint Embedding Predictive Architectures (JEPA) offer a promising approach to self-supervised speech representation learning, but suffer from representation collapse without explicit grounding. We propose GMM-Anchored JEPA, which fits a Gaussian Mixture Model once on log-mel spectrograms and uses its frozen soft posteriors as auxiliary targets throughout training. A decaying supervision schedule allows GMM regularization to dominate early training before gradually yielding to the JEPA objective. Unlike HuBERT and WavLM, which require iterative re-clustering, our approach clusters input features once with soft rather than hard assignments. On ~50k hours of speech, GMM anchoring improves ASR (28.68% vs. 33.22% WER), emotion recognition (67.76% vs. 65.46%), and slot filling (64.7% vs. 59.1% F1) compared to a WavLM-style baseline with matched compute. Cluster analysis shows GMM-anchored representations achieve up to 98% entropy compared to 31% for WavLM-style, indicating substantially more uniform cluster utilization. Code is made available at https://github.com/gioannides/clustering-anchored-jepa.",
      "authors": [
        "Georgios Ioannides",
        "Adrian Kieback",
        "Judah Goldfeder",
        "Linsey Pang",
        "Aman Chadha",
        "Aaron Elkins",
        "Yann LeCun",
        "Ravid Shwartz-Ziv"
      ],
      "published_at": "2026-01-30T20:51:37",
      "categories": [
        "eess.AS",
        "cs.AI",
        "cs.LG",
        "cs.SD"
      ],
      "url": "http://arxiv.org/abs/2602.09040v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "application",
      "domain_bucket": "audio",
      "paper_type": "systems",
      "narrow_domain_flag": false,
      "sim_public": 0.4621236324310303,
      "sim_memory": 0.42082157731056213,
      "sim_negative": 0.2856307923793793,
      "broad_relevance": 0.4621236324310303,
      "momentum": 0.7383333333333334,
      "teachability": 0.7,
      "novelty_score": 0.5271482467651367,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.42082157731056213,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.4,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5653759934107463,
      "memory_score": 0.22224647319316865,
      "quality_score": 0.49500000000000005,
      "bridge_score": 0.22224647319316865,
      "max_axis_score": 0.5653759934107463,
      "influencer_boost": 0.2375,
      "influencer_matches": [
        "Yann LeCun"
      ],
      "pwc_trending_flag": false,
      "social_score": 0.2375,
      "scoring_sources": [
        "influencer:Yann LeCun",
        "social_momentum",
        "compound_window:90d"
      ],
      "time_window": "90d",
      "compound_window_boost": 0.0315,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2512.02012",
      "title": "Improved Mean Flows: On the Challenges of Fastforward Generative Models",
      "abstract": "MeanFlow (MF) has recently been established as a framework for one-step generative modeling. However, its ``fastforward'' nature introduces key challenges in both the training objective and the guidance mechanism. First, the original MF's training target depends not only on the underlying ground-truth fields but also on the network itself. To address this issue, we recast the objective as a loss on the instantaneous velocity $v$, re-parameterized by a network that predicts the average velocity $u$. Our reformulation yields a more standard regression problem and improves the training stability. Second, the original MF fixes the classifier-free guidance scale during training, which sacrifices flexibility. We tackle this issue by formulating guidance as explicit conditioning variables, thereby retaining flexibility at test time. The diverse conditions are processed through in-context conditioning, which reduces model size and benefits performance. Overall, our $\\textbf{improved MeanFlow}$ ($\\textbf{iMF}$) method, trained entirely from scratch, achieves $\\textbf{1.72}$ FID with a single function evaluation (1-NFE) on ImageNet 256$\\times$256. iMF substantially outperforms prior methods of this kind and closes the gap with multi-step methods while using no distillation. We hope our work will further advance fastforward generative modeling as a stand-alone paradigm.",
      "authors": [
        "Zhengyang Geng",
        "Yiyang Lu",
        "Zongze Wu",
        "Eli Shechtman",
        "J. Zico Kolter",
        "Kaiming He"
      ],
      "published_at": "2025-12-01T18:59:49",
      "categories": [
        "cs.CV",
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2512.02012v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "application",
      "domain_bucket": "vision",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.4327046275138855,
      "sim_memory": 0.40673017501831055,
      "sim_negative": 0.29411280155181885,
      "broad_relevance": 0.4327046275138855,
      "momentum": 0.7833333333333333,
      "teachability": 0.6,
      "novelty_score": 0.5321260690689087,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.40673017501831055,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.5462969652811687,
      "memory_score": 0.21801905250549314,
      "quality_score": 0.525,
      "bridge_score": 0.21801905250549314,
      "max_axis_score": 0.5462969652811687,
      "influencer_boost": 0.2375,
      "influencer_matches": [
        "Kaiming He"
      ],
      "pwc_trending_flag": false,
      "social_score": 0.2375,
      "scoring_sources": [
        "influencer:Kaiming He",
        "social_momentum",
        "compound_window:180d"
      ],
      "time_window": "180d",
      "compound_window_boost": 0.045,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2603.28765",
      "title": "Adaptive Block-Scaled Data Types",
      "abstract": "NVFP4 has grown increasingly popular as a 4-bit format for quantizing large language models due to its hardware support and its ability to retain useful information with relatively few bits per parameter. However, the format is not without limitations: recent work has shown that NVFP4 suffers from its error distribution, resulting in large amounts of quantization error on near-maximal values in each group of 16 values. In this work, we leverage this insight to design new Adaptive Block-Scaled Data Types that can adapt to the distribution of their input values. For four-bit quantization, our proposed IF4 (Int/Float 4) data type selects between FP4 and INT4 representations for each group of 16 values, which are then scaled by an E4M3 scale factor as is done with NVFP4. The selected data type is denoted using the scale factor's sign bit, which is currently unused in NVFP4, and we apply the same insight to design formats for other bit-widths, including IF3 and IF6. When used to quantize language models, we find that IF4 outperforms existing 4-bit block-scaled formats, achieving lower loss during quantized training and achieving higher accuracy on many tasks in post-training quantization. We additionally design and evaluate an IF4 Multiply-Accumulate (MAC) unit to demonstrate that IF4 can be implemented efficiently in next-generation hardware accelerators. Our code is available at https://github.com/mit-han-lab/fouroversix.",
      "authors": [
        "Jack Cook",
        "Hyemin S. Lee",
        "Kathryn Le",
        "Junxian Guo",
        "Giovanni Traverso",
        "Anantha P. Chandrakasan",
        "Song Han"
      ],
      "published_at": "2026-03-30T17:59:33",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2603.28765v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.44637954235076904,
      "sim_memory": 0.7055987119674683,
      "sim_negative": 0.1395280957221985,
      "broad_relevance": 0.44637954235076904,
      "momentum": 0.6000000000000001,
      "teachability": 0.75,
      "novelty_score": 0.513735294342041,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.7055987119674683,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.5409741568565369,
      "memory_score": 0.3076796135902405,
      "quality_score": 0.525,
      "bridge_score": 0.3076796135902405,
      "max_axis_score": 0.5409741568565369,
      "influencer_boost": 0.225,
      "influencer_matches": [
        "Song Han"
      ],
      "pwc_trending_flag": false,
      "social_score": 0.225,
      "scoring_sources": [
        "influencer:Song Han",
        "social_momentum"
      ],
      "time_window": "30d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    }
  ],
  "memory": [
    {
      "arxiv_id": "2512.01541",
      "title": "RoMe: Row Granularity Access Memory System for Large Language Models",
      "abstract": "Modern HBM-based memory systems have evolved over generations while retaining cache line granularity accesses. Preserving this fine granularity necessitated the introduction of bank groups and pseudo channels. These structures expand timing parameters and control overhead, significantly increasing memory controller scheduling complexity. Large language models (LLMs) now dominate deep learning workloads, streaming contiguous data blocks ranging from several kilobytes to megabytes per operation. In a conventional HBM-based memory system, these transfers are fragmented into hundreds of 32B cache line transactions. This forces the memory controller to employ unnecessarily intricate scheduling, leading to growing inefficiency.   To address this problem, we propose RoMe. RoMe accesses DRAM at row granularity and removes columns, bank groups, and pseudo channels from the memory interface. This design simplifies memory scheduling, thereby requiring fewer pins per channel. The freed pins are aggregated to form additional channels, increasing overall bandwidth by 12.5% with minimal extra pins. RoMe demonstrates how memory scheduling logic can be significantly simplified for representative LLM workloads, and presents an alternative approach for next-generation HBM-based memory systems achieving increased bandwidth with minimal hardware overhead.",
      "authors": [
        "Hwayong Nam",
        "Seungmin Baek",
        "Jumin Kim",
        "Michael Jaemin Kim",
        "Jung Ho Ahn"
      ],
      "published_at": "2025-12-01T11:14:31",
      "categories": [
        "cs.AR"
      ],
      "url": "http://arxiv.org/abs/2512.01541v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "hardware",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.41210484504699707,
      "sim_memory": 0.6038755774497986,
      "sim_negative": 0.14623583853244781,
      "broad_relevance": 0.41210484504699707,
      "momentum": 0.0,
      "teachability": 0.7,
      "novelty_score": 0.4122152328491211,
      "evidence_score": 0.48000000000000004,
      "direct_memory_relevance": 0.6038755774497986,
      "systems_leverage": 0.5499999999999999,
      "deployment_proximity": 0.5,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.6000000000000001,
      "transferability_score": 0.62,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.34546373844146727,
      "memory_score": 0.7231626732349395,
      "quality_score": 0.545,
      "bridge_score": 0.34546373844146727,
      "max_axis_score": 0.7231626732349395,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Memory/Storage Core",
        "Systems",
        "Hardware",
        "Inference",
        "Memory-first"
      ],
      "status": "Monitor",
      "why_now": "LLM serving is exposing the mismatch between cache-line-oriented HBM interfaces and the large contiguous transfers real models perform, so a row-granularity memory design aimed at LLMs is timely. The paper directly targets bandwidth, controller complexity, and pin efficiency in a part of the stack that is becoming a real deployment bottleneck.",
      "why_not_higher": "This is a hardware-architecture proposal with limited immediate adoption path for most AI practitioners, since it depends on next-generation memory-system design rather than a software change people can deploy now. Its broad-audience appeal is therefore narrower than papers that directly change model training or serving practice this quarter.",
      "downgrade_reasons": [
        "hardware adoption path is long",
        "impact depends on future memory-system integration",
        "broad AI audience may see it as infrastructure-specialist work"
      ],
      "what_would_raise_priority": "End-to-end results on widely recognized LLM serving stacks or evidence that vendors are moving toward similar row-granularity memory interfaces would raise it.",
      "one_sentence_episode_hook": "What if the real bottleneck in LLM inference is not the model, but the fact that modern HBM still feeds giant tensor streams through tiny cache-line transactions?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2603.27624",
      "title": "Expert Streaming: Accelerating Low-Batch MoE Inference via Multi-chiplet Architecture and Dynamic Expert Trajectory Scheduling",
      "abstract": "Mixture-of-Experts is a promising approach for edge AI with low-batch inference. Yet, on-device deployments often face limited on-chip memory and severe workload imbalance; the prevalent use of offloading further incurs off-chip memory access bottlenecks. Moreover, MoE sparsity and dynamic gating shift distributed strategies toward much finer granularity and introduce runtime scheduling considerations. Recently, high die-to-die bandwidth chiplet interconnects have created new opportunities for multi-chiplet systems to address workload imbalance and offloading bottlenecks with fine-grained scheduling. In this paper, we propose Fully Sharded Expert Data Parallelism, a parallelization paradigm specifically architected for low-batch MoE inference on multi-chiplet accelerators. FSE-DP attains adaptive computation-communication overlap and balanced load by orchestrating fine-grained, complementary expert streams along dynamic trajectories across high-bandwidth D2D links. The attendant dataflow complexity is tamed by a minimal, hardware-amenable set of virtualization rules and a lightweight scheduling algorithm. Our approach achieves 1.22 to 2.00 times speedup over state-of-the-art baselines and saves up to 78.8 percent on-chip memory.",
      "authors": [
        "Songchen Ma",
        "Hongyi Li",
        "Weihao Zhang",
        "Yonghao Tan",
        "Pingcheng Dong",
        "Yu Liu",
        "Lan Liu",
        "Yuzhong Jiao",
        "Xuejiao Liu",
        "Luhong Liang"
      ],
      "published_at": "2026-03-29T10:44:55",
      "categories": [
        "cs.AR",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2603.27624v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "hardware",
      "domain_bucket": "other",
      "paper_type": "application",
      "narrow_domain_flag": false,
      "sim_public": 0.5285307168960571,
      "sim_memory": 0.5563932657241821,
      "sim_negative": 0.1867237687110901,
      "broad_relevance": 0.5285307168960571,
      "momentum": 0.0,
      "teachability": 0.7,
      "novelty_score": 0.519289642572403,
      "evidence_score": 0.25,
      "direct_memory_relevance": 0.5563932657241821,
      "systems_leverage": 0.7,
      "deployment_proximity": 0.5,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.6000000000000001,
      "transferability_score": 0.52,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.3,
      "public_interest_score": 0.4614526614546775,
      "memory_score": 0.7139179797172547,
      "quality_score": 0.47500000000000003,
      "bridge_score": 0.4614526614546775,
      "max_axis_score": 0.7139179797172547,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "30d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Memory/Storage Core",
        "Systems",
        "Hardware",
        "Inference",
        "Memory-first"
      ],
      "status": "Monitor",
      "why_now": "MoE serving is pushing memory capacity, bandwidth, and load-balancing limits, and this paper attacks those bottlenecks directly with a chiplet-oriented inference design. It is timely because sparse models keep exposing data-movement costs that standard single-die assumptions hide.",
      "why_not_higher": "The contribution is tied to low-batch MoE inference on a specific multi-chiplet architecture, so the audience is narrower than a broadly applicable serving paper. The headline gains are interesting, but the impact depends on how representative the hardware assumptions and workloads are for real deployments.",
      "downgrade_reasons": [
        "hardware-specific design",
        "focused on low-batch MoE inference",
        "transferability beyond chiplet MoE serving is limited",
        "benchmark gains may depend on architecture assumptions"
      ],
      "what_would_raise_priority": "Stronger evidence on widely used MoE models and clearer guidance for how these scheduling ideas transfer to mainstream serving stacks would raise it.",
      "one_sentence_episode_hook": "What if the real bottleneck in MoE inference is not the experts themselves, but how memory traffic and expert routes bounce across chiplets?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2603.10031",
      "title": "Architecture-Aware LLM Inference Optimization on AMD Instinct GPUs: A Comprehensive Benchmark and Deployment Study",
      "abstract": "We present a cross-architecture evaluation of production LLM inference on AMD Instinct MI325X GPUs, benchmarking four models spanning 235B to 1 trillion parameters across three architectural families (MoE+MLA, Dense+GQA, MoE+GQA) on an 8-GPU cluster with 2TB aggregate HBM3e using vLLM v0.14.1. Our results demonstrate that architecture-aware optimization is essential: MLA models require block size 1 and cannot use KV cache offloading, while GQA models benefit from both. The AMD AITER runtime is required for competitive MLA inference throughput and must be selectively disabled for architectures with incompatible attention head configurations. A controlled AITER ablation on Llama-3.1-405B (n=5 per condition) reveals a modest 3-5% throughput benefit at high concurrency but 2-16x higher measurement variability, confirming that AITER's large speedups target MoE/MLA kernels specifically. Under text-only workloads, Llama-405B and DeepSeek V3.2 achieve comparable peak throughput (15,944 and 15,343 tok/s) despite an order-of-magnitude difference in active parameters. Under vision workloads, Qwen3-VL-235B reaches 47,873 tok/s, 6.5x higher than Kimi-K2.5 (7,327 tok/s). Active parameter count per token is associated with inference throughput, though confounded by differences in quantization, AITER acceleration, and tensor parallelism. All four models exhibit a common throughput saturation point consistent with a memory-bandwidth bottleneck (~500 concurrent for short sequences, ~100-200 for longer sequences). All models maintain 100% HTTP-level success rates through 1,000 concurrent users, processing 18.9 million tokens across 17,406 requests without failures.",
      "authors": [
        "Athos Georgiou"
      ],
      "published_at": "2026-02-27T13:21:48",
      "categories": [
        "cs.AR",
        "cs.AI",
        "cs.DC"
      ],
      "url": "http://arxiv.org/abs/2603.10031v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "hardware",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.45027071237564087,
      "sim_memory": 0.5135151147842407,
      "sim_negative": 0.2471168339252472,
      "broad_relevance": 0.45027071237564087,
      "momentum": 0.0,
      "teachability": 0.6,
      "novelty_score": 0.4930337071418762,
      "evidence_score": 0.6000000000000001,
      "direct_memory_relevance": 0.5135151147842407,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.5,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.6000000000000001,
      "transferability_score": 0.52,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.46903626978397367,
      "memory_score": 0.6190545344352723,
      "quality_score": 0.525,
      "bridge_score": 0.46903626978397367,
      "max_axis_score": 0.6190545344352723,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Memory/Storage Core",
        "Bridge",
        "Systems",
        "Hardware",
        "Inference",
        "Memory-first"
      ],
      "status": "Monitor",
      "why_now": "Large-model serving is being re-optimized around memory bandwidth, KV behavior, and architecture-specific runtimes, and this paper gives concrete deployment data on AMD's newest high-memory inference hardware. It is timely because teams are actively evaluating alternatives to NVIDIA for production LLM inference.",
      "why_not_higher": "The study is strong operationally, but much of its value is tied to AMD Instinct, vLLM, and AITER-specific behavior rather than broadly portable serving guidance. It risks reading as a careful vendor/platform benchmark more than a field-shifting systems result.",
      "downgrade_reasons": [
        "Vendor-specific deployment study",
        "Transferability depends on AMD/vLLM/AITER stack",
        "Benchmark-heavy paper with limited new mechanism"
      ],
      "what_would_raise_priority": "Clearer cross-platform comparisons or reusable rules for KV offload, bandwidth saturation, and runtime selection across multiple serving stacks would raise it.",
      "one_sentence_episode_hook": "What actually bottlenecks trillion-parameter inference on AMD GPUs: model architecture, KV cache policy, or raw memory bandwidth?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2512.24511",
      "title": "Understanding LLM Checkpoint/Restore I/O Strategies and Patterns",
      "abstract": "As LLMs and foundation models scale, checkpoint/restore has become a critical pattern for training and inference. With 3D parallelism (tensor, pipeline, data), checkpointing involves many processes, each managing numerous tensors of varying shapes and sizes, that must be persisted frequently to stable storage (e.g., parallel file systems). This turns checkpoint/restore into a big-data I/O problem characterized by volume, variety, and velocity. The workflow must traverse the full storage stack -- from GPU memory through host memory and local storage to external repositories -- whose tiers differ by orders of magnitude in performance, creating bottlenecks under concurrency even with asynchronous flush/prefetch. Kernel-accelerated I/O libraries such as \\texttt{liburing} may mitigate these issues versus POSIX, but their effectiveness for LLM checkpointing remains underexplored. We develop microbenchmarks to quantify trade-offs when using \\texttt{liburing}, evaluating how aggregation, alignment, and I/O coalescing interact under buffered and direct I/O. We find that uncoalesced small-buffer operations halve throughput relative to synthetic workloads, while file system-aware aggregation restores bandwidth and reduces metadata overhead. Compared to state-of-the-art LLM checkpointing engines, our approach achieves up to $3.9\\times$ higher write throughput than DataStates-LLM and $7.6\\times$ higher than TorchSnapshot. These results highlight the need for aggregation and coalescing strategies that align with modern file systems and I/O backends.",
      "authors": [
        "Mikaila J. Gossman",
        "Avinash Maurya",
        "Bogdan Nicolae",
        "Jon C. Calhoun"
      ],
      "published_at": "2025-12-30T23:21:02",
      "categories": [
        "cs.DC"
      ],
      "url": "http://arxiv.org/abs/2512.24511v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "systems",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.5061847567558289,
      "sim_memory": 0.5454308390617371,
      "sim_negative": 0.20787817239761353,
      "broad_relevance": 0.5061847567558289,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.5080681443214417,
      "evidence_score": 0.49,
      "direct_memory_relevance": 0.5454308390617371,
      "systems_leverage": 0.5499999999999999,
      "deployment_proximity": 0.7,
      "memory_adjacent_future_value": 0.25,
      "bandwidth_capacity": 0.2,
      "transferability_score": 0.6699999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.3980656486749649,
      "memory_score": 0.6306292517185211,
      "quality_score": 0.525,
      "bridge_score": 0.3980656486749649,
      "max_axis_score": 0.6306292517185211,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Memory/Storage Core",
        "Systems",
        "Training",
        "Memory-first"
      ],
      "status": "Monitor",
      "why_now": "Checkpoint and restore are turning into real scaling bottlenecks as large-model training and recovery workflows push more state through mismatched storage tiers. This paper matters now because it connects LLM systems practice to concrete I/O strategy choices that can materially change throughput.",
      "why_not_higher": "The topic is important but still fairly infrastructure-specific, and the paper reads more like a careful systems benchmark than a shift most AI builders will act on immediately. Its strongest results are about checkpoint write-path engineering rather than a broadly adopted new training or serving method.",
      "downgrade_reasons": [
        "Broad audience appeal is limited outside training systems and HPC operators",
        "Benchmark-heavy framing makes episode narrative less naturally dramatic",
        "Impact is clearer for checkpoint infrastructure than for everyday model building or serving"
      ],
      "what_would_raise_priority": "Evidence that these I/O strategies improve end-to-end training wall-clock, failure recovery, or production-scale checkpoint cadence on widely used LLM stacks would raise it.",
      "one_sentence_episode_hook": "Why are giant LLM checkpoints still crawling through the storage stack, and what happens when smarter I/O coalescing suddenly makes state saves several times faster?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2603.27138",
      "title": "ScoutAttention: Efficient KV Cache Offloading via Layer-Ahead CPU Pre-computation for LLM Inference",
      "abstract": "Large language models encounter critical GPU memory capacity constraints during long-context inference, where KV cache memory consumption severely limits decode batch sizes. While existing research has explored offloading KV cache to DRAM, these approaches either demand frequent GPU-CPU data transfers or impose extensive CPU computation requirements, resulting in poor GPU utilization as the system waits for I/O operations or CPU processing to complete.   We propose ScoutAttention, a novel KV cache offloading framework that accelerates LLM inference through collaborative GPU-CPU attention computation. To prevent CPU computation from bottlenecking the system, ScoutAttention introduces GPU-CPU collaborative block-wise sparse attention that significantly reduces CPU load. Unlike conventional parallel computing approaches, our framework features a novel layer-ahead CPU pre-computation algorithm, enabling the CPU to initiate attention computation one layer in advance, complemented by asynchronous periodic recall mechanisms to maintain minimal CPU compute load. Experimental results demonstrate that ScoutAttention maintains accuracy within 2.4% of baseline while achieving 2.1x speedup compared to existing offloading methods.",
      "authors": [
        "Qiuyang Zhang",
        "Kai Zhou",
        "Ding Tang",
        "Kai Lu",
        "Cheng Li",
        "Zhenyu Yang",
        "Peng Xu",
        "Jiguang Wan"
      ],
      "published_at": "2026-03-28T05:06:05",
      "categories": [
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2603.27138v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.4604194760322571,
      "sim_memory": 0.6243070363998413,
      "sim_negative": 0.21835364401340485,
      "broad_relevance": 0.4604194760322571,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.45993274450302124,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.6243070363998413,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.6000000000000001,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.43711575448513035,
      "memory_score": 0.4002921109199524,
      "quality_score": 0.595,
      "bridge_score": 0.4002921109199524,
      "max_axis_score": 0.43711575448513035,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "30d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Memory-first"
      ],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2512.02278",
      "title": "Fantasy: Efficient Large-scale Vector Search on GPU Clusters with GPUDirect Async",
      "abstract": "Vector similarity search has become a critical component in AI-driven applications such as large language models (LLMs). To achieve high recall and low latency, GPUs are utilized to exploit massive parallelism for faster query processing. However, as the number of vectors continues to grow, the graph size quickly exceeds the memory capacity of a single GPU, making it infeasible to store and process the entire index on a single GPU. Recent work uses CPU-GPU architectures to keep vectors in CPU memory or SSDs, but the loading step stalls GPU computation. We present Fantasy, an efficient system that pipelines vector search and data transfer in a GPU cluster with GPUDirect Async. Fantasy overlaps computation and network communication to significantly improve search throughput for large graphs and deliver large query batch sizes.",
      "authors": [
        "Yi Liu",
        "Chen Qian"
      ],
      "published_at": "2025-12-01T23:47:51",
      "categories": [
        "cs.DC"
      ],
      "url": "http://arxiv.org/abs/2512.02278v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "systems",
      "domain_bucket": "llm",
      "paper_type": "systems",
      "narrow_domain_flag": false,
      "sim_public": 0.41231590509414673,
      "sim_memory": 0.44079989194869995,
      "sim_negative": 0.17094996571540833,
      "broad_relevance": 0.41231590509414673,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.4705544114112854,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.44079989194869995,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.7,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.4,
      "transferability_score": 0.6,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.4042779332399369,
      "memory_score": 0.41723996758460996,
      "quality_score": 0.525,
      "bridge_score": 0.4042779332399369,
      "max_axis_score": 0.41723996758460996,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Memory-first"
      ],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2512.22838",
      "title": "OrchANN: A Unified I/O Orchestration Framework for Skewed Out-of-Core Vector Search",
      "abstract": "Approximate nearest neighbor search (ANNS) at billion scale is fundamentally an out-of-core problem: vectors and indexes live on SSD, so performance is dominated by I/O rather than compute. Under skewed semantic embeddings, existing out-of-core systems break down: a uniform local index mismatches cluster scales, static routing misguides queries and inflates the number of probed partitions, and pruning is incomplete at the cluster level and lossy at the vector level, triggering \"fetch-to-discard\" reranking on raw vectors.   We present OrchANN, an out-of-core ANNS engine that uses an I/O orchestration model for unified I/O governance along the route-access-verify pipeline. OrchANN selects a heterogeneous local index per cluster via offline auto-profiling, maintains a query-aware in-memory navigation graph that adapts to skewed workloads, and applies multi-level pruning with geometric bounds to filter both clusters and vectors before issuing SSD reads. Across five standard datasets under strict out-of-core constraints, OrchANN outperforms four baselines including DiskANN, Starling, SPANN, and PipeANN in both QPS and latency while reducing SSD accesses. Furthermore, OrchANN delivers up to 17.2x higher QPS and 25.0x lower latency than competing systems without sacrificing accuracy.",
      "authors": [
        "Chengying Huan",
        "Lizheng Chen",
        "Zhengyi Yang",
        "Shaonan Ma",
        "Rong Gu",
        "Renjie Yao",
        "Zhibin Wang",
        "Mingxing Zhang",
        "Fang Xi",
        "Jie Tao"
      ],
      "published_at": "2025-12-28T08:42:38",
      "categories": [
        "cs.DB",
        "cs.IR"
      ],
      "url": "http://arxiv.org/abs/2512.22838v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "systems",
      "domain_bucket": "recommendation",
      "paper_type": "systems",
      "narrow_domain_flag": true,
      "sim_public": 0.4004133343696594,
      "sim_memory": 0.4831111431121826,
      "sim_negative": 0.1828860640525818,
      "broad_relevance": 0.4004133343696594,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.5926013588905334,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.4831111431121826,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.7,
      "memory_adjacent_future_value": 0.25,
      "bandwidth_capacity": 0.2,
      "transferability_score": 0.5,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.4190142041444778,
      "memory_score": 0.4249333429336548,
      "quality_score": 0.5,
      "bridge_score": 0.4190142041444778,
      "max_axis_score": 0.4249333429336548,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Memory-first"
      ],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2511.00603",
      "title": "EPARA: Parallelizing Categorized AI Inference in Edge Clouds",
      "abstract": "With the increasing adoption of AI applications such as large language models and computer vision AI, the computational demands on AI inference systems are continuously rising, making the enhancement of task processing capacity using existing hardware a primary objective in edge clouds. We propose EPARA, an end-to-end AI parallel inference framework in edge, aimed at enhancing the edge AI serving capability. Our key idea is to categorize tasks based on their sensitivity to latency/frequency and requirement for GPU resources, thereby achieving both request-level and service-level task-resource allocation. EPARA consists of three core components: 1) a task-categorized parallelism allocator that decides the parallel mode of each task, 2) a distributed request handler that performs the calculation for the specific request, and 3) a state-aware scheduler that periodically updates service placement in edge clouds. We implement a EPARA prototype and conduct a case study on the EPARA operation for LLMs and segmentation tasks. Evaluation through testbed experiments involving edge servers, embedded devices, and microcomputers shows that EPARA achieves up to 2.1$\\times$ higher goodput in production workloads compared to prior frameworks, while adapting to various edge AI inference tasks.",
      "authors": [
        "Yubo Wang",
        "Yubo Cui",
        "Tuo Shi",
        "Danyang Li",
        "Wenxin Li",
        "Lide Suo",
        "Tao Wang",
        "Xin Xie"
      ],
      "published_at": "2025-11-01T16:09:14",
      "categories": [
        "cs.DC",
        "cs.AI",
        "cs.NI"
      ],
      "url": "http://arxiv.org/abs/2511.00603v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "systems",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.47736793756484985,
      "sim_memory": 0.49913936853408813,
      "sim_negative": 0.3375422954559326,
      "broad_relevance": 0.47736793756484985,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.6284320950508118,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.49913936853408813,
      "systems_leverage": 0.5499999999999999,
      "deployment_proximity": 0.7,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.4674751955270768,
      "memory_score": 0.4017418105602264,
      "quality_score": 0.595,
      "bridge_score": 0.4017418105602264,
      "max_axis_score": 0.4674751955270768,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Memory-first"
      ],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2512.24157",
      "title": "Training Report of TeleChat3-MoE",
      "abstract": "TeleChat3-MoE is the latest series of TeleChat large language models, featuring a Mixture-of-Experts (MoE) architecture with parameter counts ranging from 105 billion to over one trillion,trained end-to-end on Ascend NPU cluster. This technical report mainly presents the underlying training infrastructure that enables reliable and efficient scaling to frontier model sizes. We detail systematic methodologies for operator-level and end-to-end numerical accuracy verification, ensuring consistency across hardware platforms and distributed parallelism strategies. Furthermore, we introduce a suite of performance optimizations, including interleaved pipeline scheduling, attention-aware data scheduling for long-sequence training,hierarchical and overlapped communication for expert parallelism, and DVM-based operator fusion. A systematic parallelization framework, leveraging analytical estimation and integer linear programming, is also proposed to optimize multi-dimensional parallelism configurations. Additionally, we present methodological approaches to cluster-level optimizations, addressing host- and device-bound bottlenecks during large-scale training tasks. These infrastructure advancements yield significant throughput improvements and near-linear scaling on clusters comprising thousands of devices, providing a robust foundation for large-scale language model development on hardware ecosystems.",
      "authors": [
        "Xinzhang Liu",
        "Chao Wang",
        "Zhihao Yang",
        "Zhuo Jiang",
        "Xuncheng Zhao",
        "Haoran Wang",
        "Lei Li",
        "Dongdong He",
        "Luobin Liu",
        "Kaizhe Yuan"
      ],
      "published_at": "2025-12-30T11:42:14",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2512.24157v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "systems",
      "narrow_domain_flag": false,
      "sim_public": 0.6095997095108032,
      "sim_memory": 0.540129542350769,
      "sim_negative": 0.27689701318740845,
      "broad_relevance": 0.6095997095108032,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.5845089256763458,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.540129542350769,
      "systems_leverage": 0.85,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.48055625170469285,
      "memory_score": 0.39303886270523075,
      "quality_score": 0.575,
      "bridge_score": 0.39303886270523075,
      "max_axis_score": 0.48055625170469285,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Memory-relevant"
      ],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2603.28290",
      "title": "OptINC: Optical In-Network-Computing for Scalable Distributed Learning",
      "abstract": "Distributed learning is widely used for training large models on large datasets by distributing parts of the model or dataset across multiple devices and aggregating the computed results for subsequent computations or parameter updates. Existing communication algorithms for distributed learning such as ring all-reduce result in heavy communication overhead between servers. Since communication in large-scale systems uses optical fibers, we propose an Optical In-Network-Computing (OptINC) architecture to offload the computation in servers onto the optical interconnects. To execute gradient averaging and quantization in the optical domain, we incorporate optical devices such as Mach-Zehnder-Interferometers (MZIs) into the interconnects. Such a de facto optical neural network (ONN) can effectively reduce the communication overhead in existing distributed training solutions. To reduce dataset complexity for training this neural network, a preprocessing algorithm implemented in the optical domain is also proposed. Hardware cost is lowered by approximating the weight matrices of the optical neural network with unitary and diagonal matrices, while the accuracy is maintained by a proposed hardware-aware training algorithm. The proposed solution was evaluated on real distributed learning tasks, including ResNet50 on CIFAR-100, and a LLaMA-based network on Wikipedia-1B. In both cases, the proposed framework can achieve comparable training accuracy to the ring all-reduce baseline, while eliminating communication overhead.",
      "authors": [
        "Sijie Fei",
        "Grace Li Zhang",
        "Bing Li",
        "Ulf Schlichtmann"
      ],
      "published_at": "2026-03-30T11:10:47",
      "categories": [
        "cs.LG",
        "cs.AR"
      ],
      "url": "http://arxiv.org/abs/2603.28290v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.549875020980835,
      "sim_memory": 0.4452601373195648,
      "sim_negative": 0.3170567750930786,
      "broad_relevance": 0.549875020980835,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.5667058825492859,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.4452601373195648,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.4,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.4799683886766434,
      "memory_score": 0.34357804119586943,
      "quality_score": 0.595,
      "bridge_score": 0.34357804119586943,
      "max_axis_score": 0.4799683886766434,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "30d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Memory-relevant"
      ],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    }
  ],
  "monitor": [
    {
      "arxiv_id": "2603.01264",
      "title": "S2O: Enhancing Adversarial Training with Second-Order Statistics of Weights",
      "abstract": "Adversarial training has emerged as a highly effective way to improve the robustness of deep neural networks (DNNs). It is typically conceptualized as a min-max optimization problem over model weights and adversarial perturbations, where the weights are optimized using gradient descent methods, such as SGD. In this paper, we propose a novel approach by treating model weights as random variables, which paves the way for enhancing adversarial training through \\textbf{S}econd-Order \\textbf{S}tatistics \\textbf{O}ptimization (S$^2$O) over model weights. We challenge and relax a prevalent, yet often unrealistic, assumption in prior PAC-Bayesian frameworks: the statistical independence of weights. From this relaxation, we derive an improved PAC-Bayesian robust generalization bound. Our theoretical developments suggest that optimizing the second-order statistics of weights can substantially tighten this bound. We complement this theoretical insight by conducting an extensive set of experiments that demonstrate that S$^2$O not only enhances the robustness and generalization of neural networks when used in isolation, but also seamlessly augments other state-of-the-art adversarial training techniques. The code is available at https://github.com/Alexkael/S2O.",
      "authors": [
        "Gaojie Jin",
        "Xinping Yi",
        "Wei Huang",
        "Sven Schewe",
        "Xiaowei Huang"
      ],
      "published_at": "2026-03-01T20:43:16",
      "categories": [
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2603.01264v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 3,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.42938750982284546,
      "sim_memory": 0.4066287875175476,
      "sim_negative": 0.21473261713981628,
      "broad_relevance": 0.42938750982284546,
      "momentum": 0.04279025907815965,
      "teachability": 0.85,
      "novelty_score": 0.5430323481559753,
      "evidence_score": 0.538629436111989,
      "direct_memory_relevance": 0.4066287875175476,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.46962357240268027,
      "memory_score": 0.23462416858870294,
      "quality_score": 0.6504517744447955,
      "bridge_score": 0.23462416858870294,
      "max_axis_score": 0.46962357240268027,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "30d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2603.00730",
      "title": "MO-MIX: Multi-Objective Multi-Agent Cooperative Decision-Making With Deep Reinforcement Learning",
      "abstract": "Deep reinforcement learning (RL) has been applied extensively to solve complex decision-making problems. In many real-world scenarios, tasks often have several conflicting objectives and may require multiple agents to cooperate, which are the multi-objective multi-agent decision-making problems. However, only few works have been conducted on this intersection. Existing approaches are limited to separate fields and can only handle multi-agent decision-making with a single objective, or multi-objective decision-making with a single agent. In this paper, we propose MO-MIX to solve the multi-objective multi-agent reinforcement learning (MOMARL) problem. Our approach is based on the centralized training with decentralized execution (CTDE) framework. A weight vector representing preference over the objectives is fed into the decentralized agent network as a condition for local action-value function estimation, while a mixing network with parallel architecture is used to estimate the joint action-value function. In addition, an exploration guide approach is applied to improve the uniformity of the final non-dominated solutions. Experiments demonstrate that the proposed method can effectively solve the multi-objective multi-agent cooperative decision-making problem and generate an approximation of the Pareto set. Our approach not only significantly outperforms the baseline method in all four kinds of evaluation metrics, but also requires less computational cost.",
      "authors": [
        "Tianmeng Hu",
        "Biao Luo",
        "Chunhua Yang",
        "Tingwen Huang"
      ],
      "published_at": "2026-02-28T16:25:22",
      "categories": [
        "cs.AI",
        "cs.LG",
        "cs.MA"
      ],
      "url": "http://arxiv.org/abs/2603.00730v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 52,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "application",
      "narrow_domain_flag": false,
      "sim_public": 0.37483441829681396,
      "sim_memory": 0.26612287759780884,
      "sim_negative": 0.20876100659370422,
      "broad_relevance": 0.37483441829681396,
      "momentum": 0.29229351302416656,
      "teachability": 0.85,
      "novelty_score": 0.4396376609802246,
      "evidence_score": 0.6,
      "direct_memory_relevance": 0.26612287759780884,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.3,
      "public_interest_score": 0.49685467724091115,
      "memory_score": 0.19983686327934264,
      "quality_score": 0.645,
      "bridge_score": 0.19983686327934264,
      "max_axis_score": 0.49685467724091115,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [
        "compound_window:90d"
      ],
      "time_window": "90d",
      "compound_window_boost": 0.0315,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2602.24195",
      "title": "Uncertainty Quantification for Multimodal Large Language Models with Incoherence-adjusted Semantic Volume",
      "abstract": "Despite their capabilities, Multimodal Large Language Models (MLLMs) may produce plausible but erroneous outputs, hindering reliable deployment. Accurate uncertainty metrics could enable escalation of unreliable queries to human experts or larger models for improved performance. However, existing uncertainty metrics have practical constraints, such as being designed only for specific modalities, reliant on external tools, or computationally expensive. We introduce UMPIRE, a training-free uncertainty quantification framework for MLLMs that works efficiently across various input and output modalities without external tools, relying only on the models' own internal modality features. UMPIRE computes the incoherence-adjusted semantic volume of sampled MLLM responses for a given task instance, effectively capturing both the global semantic diversity of samples and the local incoherence of responses based on internal model confidence. We propose uncertainty desiderata for MLLMs and provide theoretical analysis motivating UMPIRE's design. Extensive experiments show that UMPIRE consistently outperforms baseline metrics in error detection and uncertainty calibration across image, audio, and video-text benchmarks, including adversarial and out-of-distribution settings. We also demonstrate UMPIRE's generalization to non-text output tasks, including image and audio generation.",
      "authors": [
        "Gregory Kang Ruey Lau",
        "Hieu Dao",
        "Nicole Kan Hui Lin",
        "Bryan Kian Hsiang Low"
      ],
      "published_at": "2026-02-27T17:18:42",
      "categories": [
        "cs.AI",
        "cs.CL",
        "cs.CV",
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2602.24195v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.4659190773963928,
      "sim_memory": 0.34304171800613403,
      "sim_negative": 0.20153295993804932,
      "broad_relevance": 0.4659190773963928,
      "momentum": 0.017490950964499406,
      "teachability": 0.85,
      "novelty_score": 0.6151334047317505,
      "evidence_score": 0.46931471805599456,
      "direct_memory_relevance": 0.34304171800613403,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.47594113182997955,
      "memory_score": 0.20723028156855955,
      "quality_score": 0.6227258872223977,
      "bridge_score": 0.20723028156855955,
      "max_axis_score": 0.47594113182997955,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2602.23827",
      "title": "FedNSAM:Consistency of Local and Global Flatness for Federated Learning",
      "abstract": "In federated learning (FL), multi-step local updates and data heterogeneity usually lead to sharper global minima, which degrades the performance of the global model. Popular FL algorithms integrate sharpness-aware minimization (SAM) into local training to address this issue. However, in the high data heterogeneity setting, the flatness in local training does not imply the flatness of the global model. Therefore, minimizing the sharpness of the local loss surfaces on the client data does not enable the effectiveness of SAM in FL to improve the generalization ability of the global model. We define the \\textbf{flatness distance} to explain this phenomenon. By rethinking the SAM in FL and theoretically analyzing the \\textbf{flatness distance}, we propose a novel \\textbf{FedNSAM} algorithm that accelerates the SAM algorithm by introducing global Nesterov momentum into the local update to harmonize the consistency of global and local flatness. \\textbf{FedNSAM} uses the global Nesterov momentum as the direction of local estimation of client global perturbations and extrapolation. Theoretically, we prove a tighter convergence bound than FedSAM by Nesterov extrapolation. Empirically, we conduct comprehensive experiments on CNN and Transformer models to verify the superior performance and efficiency of \\textbf{FedNSAM}. The code is available at https://github.com/junkangLiu0/FedNSAM.",
      "authors": [
        "Junkang Liu",
        "Fanhua Shang",
        "Yuxuan Tian",
        "Hongying Liu",
        "Yuanyuan Liu"
      ],
      "published_at": "2026-02-27T09:07:47",
      "categories": [
        "cs.LG",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2602.23827v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 14,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "theory",
      "narrow_domain_flag": false,
      "sim_public": 0.40056392550468445,
      "sim_memory": 0.34842756390571594,
      "sim_negative": 0.2246229648590088,
      "broad_relevance": 0.40056392550468445,
      "momentum": 0.2149105750391588,
      "teachability": 0.85,
      "novelty_score": 0.5550581216812134,
      "evidence_score": 0.570805020110221,
      "direct_memory_relevance": 0.34842756390571594,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7,
      "reproducibility": 0.3,
      "public_interest_score": 0.5020307639279523,
      "memory_score": 0.2210248715849413,
      "quality_score": 0.6133220080440884,
      "bridge_score": 0.2210248715849413,
      "max_axis_score": 0.5020307639279523,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [
        "compound_window:90d"
      ],
      "time_window": "90d",
      "compound_window_boost": 0.0315,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2603.00498",
      "title": "Antibody: Strengthening Defense Against Harmful Fine-Tuning for Large Language Models via Attenuating Harmful Gradient Influence",
      "abstract": "Fine-tuning-as-a-service introduces a threat to Large Language Models' safety when service providers fine-tune their models on poisoned user-submitted datasets, a process known as harmful fine-tuning attacks. In this work, we show that by regularizing the gradient contribution of harmful samples encountered during fine-tuning, we can effectively mitigate the impact of harmful fine-tuning attacks. To this end, we introduce Antibody, a defense strategy that first ensures robust safety alignment for the model before fine-tuning, and then applies a safety-preservation learning algorithm during fine-tuning. Specifically, in the alignment stage before fine-tuning, we propose optimizing the model to be in a flat loss region with respect to harmful samples, which makes the safety alignment more resilient to subsequent harmful fine-tuning. Then, in the fine-tuning stage, we design a fine-tuning algorithm that applies a weighting scheme to all samples in each training batch to inhibit the model from learning from harmful samples while encouraging learning from benign samples. Experimental results demonstrate that Antibody successfully mitigates harmful fine-tuning attacks while boosting fine-tuning performance on the user-submitted dataset.",
      "authors": [
        "Quoc Minh Nguyen",
        "Trung Le",
        "Jing Wu",
        "Anh Tuan Bui",
        "Mehrtash Harandi"
      ],
      "published_at": "2026-02-28T06:46:21",
      "categories": [
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2603.00498v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 3,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.47394734621047974,
      "sim_memory": 0.37424540519714355,
      "sim_negative": 0.15203379094600677,
      "broad_relevance": 0.47394734621047974,
      "momentum": 0.04279025907815965,
      "teachability": 0.85,
      "novelty_score": 0.5816698670387268,
      "evidence_score": 0.538629436111989,
      "direct_memory_relevance": 0.37424540519714355,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.48878715115138327,
      "memory_score": 0.22490915389258176,
      "quality_score": 0.6004517744447956,
      "bridge_score": 0.22490915389258176,
      "max_axis_score": 0.48878715115138327,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2603.27481",
      "title": "On Token's Dilemma: Dynamic MoE with Drift-Aware Token Assignment for Continual Learning of Large Vision Language Models",
      "abstract": "Multimodal Continual Instruction Tuning aims to continually enhance Large Vision Language Models (LVLMs) by learning from new data without forgetting previously acquired knowledge. Mixture of Experts (MoE) architectures naturally facilitate this by incrementally adding new experts and expanding routers while keeping the existing ones frozen. However, despite expert isolation, MoE-based continual learners still suffer from forgetting due to routing-drift: old-task tokens become mistakenly attracted to newly added experts, degrading performance on prior tasks. We analyze the failure mode at the token level and reveal the token's dilemma: ambiguous and old tokens in new-task data offer minimal learning benefit yet induce forgetting when routed to new experts, due to their ambiguous routing assignment during training. Motivated by this, we propose LLaVA-DyMoE, a dynamic MoE framework that incrementally expands the MoE with drift-aware token assignment. We characterize token types via their routing score distributions and apply targeted regularization. Specifically, a token-level assignment guidance steers ambiguous and old tokens away from new experts to preserve established routing patterns and alleviate routing-drift, while complementary routing score regularizations enforce expert-group separation and promote new-expert specialization. Extensive experiments demonstrate that our LLaVA-DyMoE effectively mitigates routing-drift-induced forgetting, achieving over a 7% gain in mean final accuracy and a 12% reduction in forgetting compared to baselines. The project page is https://zhaoc5.github.io/DyMoE.",
      "authors": [
        "Chongyang Zhao",
        "Mingsong Li",
        "Haodong Lu",
        "Dong Gong"
      ],
      "published_at": "2026-03-29T02:30:55",
      "categories": [
        "cs.LG",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2603.27481v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": true,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.45524200797080994,
      "sim_memory": 0.49153757095336914,
      "sim_negative": 0.2781209349632263,
      "broad_relevance": 0.45524200797080994,
      "momentum": 0.4,
      "teachability": 0.85,
      "novelty_score": 0.5127282440662384,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.49153757095336914,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.25,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5234818390011788,
      "memory_score": 0.26846127128601077,
      "quality_score": 0.595,
      "bridge_score": 0.26846127128601077,
      "max_axis_score": 0.5234818390011788,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "30d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.23180",
      "title": "TriSpec: Ternary Speculative Decoding via Lightweight Proxy Verification",
      "abstract": "Inference efficiency in Large Language Models (LLMs) is fundamentally limited by their serial, autoregressive generation, especially as reasoning becomes a key capability and response sequences grow longer. Speculative decoding (SD) offers a powerful solution, providing significant speed-ups through its lightweight drafting and parallel verification mechanism. While existing work has nearly saturated improvements in draft effectiveness and efficiency, this paper advances SD from a new yet critical perspective: the verification cost. We propose TriSpec, a novel ternary SD framework that, at its core, introduces a lightweight proxy to significantly reduce computational cost by approving easily verifiable draft sequences and engaging the full target model only when encountering uncertain tokens. TriSpec can be integrated with state-of-the-art SD methods like EAGLE-3 to further reduce verification costs, achieving greater acceleration. Extensive experiments on the Qwen3 and DeepSeek-R1-Distill-Qwen/LLaMA families show that TriSpec achieves up to 35\\% speedup over standard SD, with up to 50\\% fewer target model invocations while maintaining comparable accuracy.",
      "authors": [
        "Haoyun Jiang",
        "Junqi He",
        "Feng Hong",
        "Xinlong Yang",
        "Jianwei Zhang",
        "Zheng Li",
        "Zhengyang Zhuge",
        "Zhiyong Chen",
        "Bo Han",
        "Junyang Lin"
      ],
      "published_at": "2026-01-30T17:04:18",
      "categories": [
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.23180v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.7044159173965454,
      "sim_memory": 0.49530547857284546,
      "sim_negative": 0.20457014441490173,
      "broad_relevance": 0.7044159173965454,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.5147088468074799,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.49530547857284546,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.25,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5185311022400856,
      "memory_score": 0.26959164357185367,
      "quality_score": 0.595,
      "bridge_score": 0.26959164357185367,
      "max_axis_score": 0.5185311022400856,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2602.23881",
      "title": "LK Losses: Direct Acceptance Rate Optimization for Speculative Decoding",
      "abstract": "Speculative decoding accelerates autoregressive large language model (LLM) inference by using a lightweight draft model to propose candidate tokens that are then verified in parallel by the target model. The speedup is significantly determined by the acceptance rate, yet standard training minimizes Kullback-Leibler (KL) divergence as a proxy objective. While KL divergence and acceptance rate share the same global optimum, small draft models, having limited capacity, typically converge to suboptimal solutions where minimizing KL does not guarantee maximizing acceptance rate. To address this issue, we propose LK losses, special training objectives that directly target acceptance rate. Comprehensive experiments across four draft architectures and six target models, ranging from 8B to 685B parameters, demonstrate consistent improvements in acceptance metrics across all configurations compared to the standard KL-based training. We evaluate our approach on general, coding and math domains and report gains of up to 8-10% in average acceptance length. LK losses are easy to implement, introduce no computational overhead and can be directly integrated into any existing speculator training framework, making them a compelling alternative to the existing draft training objectives.",
      "authors": [
        "Alexander Samarin",
        "Sergei Krutikov",
        "Anton Shevtsov",
        "Sergei Skvortsov",
        "Filipp Fisin",
        "Alexander Golubev"
      ],
      "published_at": "2026-02-27T10:20:11",
      "categories": [
        "cs.LG",
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2602.23881v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.6818745136260986,
      "sim_memory": 0.5274909734725952,
      "sim_negative": 0.2754533290863037,
      "broad_relevance": 0.6818745136260986,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.4997122287750244,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.5274909734725952,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.25,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5095191884040833,
      "memory_score": 0.27924729204177856,
      "quality_score": 0.595,
      "bridge_score": 0.27924729204177856,
      "max_axis_score": 0.5095191884040833,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2603.00573",
      "title": "CoMoL: Efficient Mixture of LoRA Experts via Dynamic Core Space Merging",
      "abstract": "Large language models (LLMs) achieve remarkable performance on diverse downstream and domain-specific tasks via parameter-efficient fine-tuning (PEFT). However, existing PEFT methods, particularly MoE-LoRA architectures, suffer from limited parameter efficiency and coarse-grained adaptation due to the proliferation of LoRA experts and instance-level routing. To address these issues, we propose Core Space Mixture of LoRA (\\textbf{CoMoL}), a novel MoE-LoRA framework that incorporates expert diversity, parameter efficiency, and fine-grained adaptation. Specifically, CoMoL introduces two key components: core space experts and core space routing. Core space experts store each expert in a compact core matrix, preserving diversity while controlling parameter growth. Core space routing dynamically selects and activates the appropriate core experts for each token, enabling fine-grained, input-adaptive routing. Activated core experts are then merged via a soft-merging strategy into a single core expert, which is combined with a shared LoRA to form a specialized LoRA module. Besides, the routing network is projected into the same low-rank space as the LoRA matrices, further reducing parameter overhead without compromising expressiveness. Extensive experiments demonstrate that CoMoL retains the adaptability of MoE-LoRA architectures while achieving parameter efficiency comparable to standard LoRA, consistently outperforming existing methods across multiple tasks.",
      "authors": [
        "Jie Cao",
        "Zhenxuan Fan",
        "Zhuonan Wang",
        "Tianwei Lin",
        "Ziyuan Zhao",
        "Rolan Yan",
        "Wenqiao Zhang",
        "Feifei Shao",
        "Hongwei Wang",
        "Jun Xiao"
      ],
      "published_at": "2026-02-28T09:40:11",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2603.00573v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.6521914005279541,
      "sim_memory": 0.4712175726890564,
      "sim_negative": 0.24282509088516235,
      "broad_relevance": 0.6521914005279541,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.5185310244560242,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.4712175726890564,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.25,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5034370738267898,
      "memory_score": 0.28936527180671695,
      "quality_score": 0.595,
      "bridge_score": 0.28936527180671695,
      "max_axis_score": 0.5034370738267898,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2603.28458",
      "title": "HISA: Efficient Hierarchical Indexing for Fine-Grained Sparse Attention",
      "abstract": "Token-level sparse attention mechanisms, exemplified by DeepSeek Sparse Attention (DSA), achieve fine-grained key selection by scoring every historical token for each query using a lightweight indexer, and then computing attention only over the selected subset. While the downstream sparse attention scales efficiently, the indexer still scans the entire prefix for every query, introducing an O($L^2$) per-layer bottleneck that becomes prohibitive as context length grows. We propose HISA (Hierarchical Indexed Sparse Attention), a drop-in replacement for the indexer that transforms the search process from a flat token scan into a two-stage hierarchical procedure. First, a block-level coarse filter scores pooled block representatives to prune irrelevant regions. Then, a token-level refinement applies the original indexer only within the remaining candidate blocks. HISA preserves the exact token-level top-k sparsity pattern required by the downstream Sparse MLA operator and requires no additional training. On kernel-level benchmarks, HISA achieves a 2$\\times$ speedup at 32K context length and 4$\\times$ at 128K. On Needle-in-a-Haystack and LongBench, we directly replace the indexer in DeepSeek-V3.2 with HISA, without any fine-tuning. HISA closely matches the original DSA in quality while significantly outperforming block-sparse baselines. Moreover, the token selection sets produced by HISA and the original DSA exhibit a mean IoU greater than 99%, indicating that the efficiency gains come with virtually no impact on selection fidelity.",
      "authors": [
        "Yufei Xu",
        "Fanxu Meng",
        "Fan Jiang",
        "Yuxuan Wang",
        "Ruijie Zhou",
        "Jiexi Wu",
        "Zhixin Pan",
        "Zhaohui Wang",
        "Xiaojuan Tang",
        "Wenjie Pei"
      ],
      "published_at": "2026-03-30T13:59:51",
      "categories": [
        "cs.LG",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2603.28458v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": true,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.47141534090042114,
      "sim_memory": 0.46018508076667786,
      "sim_negative": 0.2911456525325775,
      "broad_relevance": 0.47141534090042114,
      "momentum": 0.4,
      "teachability": 0.85,
      "novelty_score": 0.4139394760131836,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.46018508076667786,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5135155236721038,
      "memory_score": 0.28805552423000336,
      "quality_score": 0.595,
      "bridge_score": 0.28805552423000336,
      "max_axis_score": 0.5135155236721038,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "30d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2603.28590",
      "title": "MonitorBench: A Comprehensive Benchmark for Chain-of-Thought Monitorability in Large Language Models",
      "abstract": "Large language models (LLMs) can generate chains of thought (CoTs) that are not always causally responsible for their final outputs. When such a mismatch occurs, the CoT no longer faithfully reflects the decision-critical factors driving the model's behavior, leading to the reduced CoT monitorability problem. However, a comprehensive and fully open-source benchmark for studying CoT monitorability remains lacking. To address this gap, we propose MonitorBench, a systematic benchmark for evaluating CoT monitorability in LLMs. MonitorBench provides: (1) a diverse set of 1,514 test instances with carefully designed decision-critical factors across 19 tasks spanning 7 categories to characterize when CoTs can be used to monitor the factors driving LLM behavior; and (2) two stress-test settings to quantify the extent to which CoT monitorability can be degraded. Extensive experiments across multiple popular LLMs with varying capabilities show that CoT monitorability is higher when producing the final target response requires structural reasoning through the decision-critical factor. Closed-source LLMs generally show lower monitorability, and there exists a negative relationship between monitorability and model capability. Moreover, both open- and closed-source LLMs can intentionally reduce monitorability under stress-tests, with monitorability dropping by up to 30% in some tasks that do not require structural reasoning over the decision-critical factors. Beyond these empirical insights, MonitorBench provides a basis for further research on evaluating future LLMs, studying advanced stress-test monitorability techniques, and developing new monitoring approaches.",
      "authors": [
        "Han Wang",
        "Yifan Sun",
        "Brian Ko",
        "Mann Talati",
        "Jiawen Gong",
        "Zimeng Li",
        "Naicheng Yu",
        "Xucheng Yu",
        "Wei Shen",
        "Vedant Jolly"
      ],
      "published_at": "2026-03-30T15:37:42",
      "categories": [
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2603.28590v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.5970848798751831,
      "sim_memory": 0.4702179431915283,
      "sim_negative": 0.1932281106710434,
      "broad_relevance": 0.5970848798751831,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.5440629720687866,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.4702179431915283,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.49073490977287293,
      "memory_score": 0.2640653829574585,
      "quality_score": 0.595,
      "bridge_score": 0.2640653829574585,
      "max_axis_score": 0.49073490977287293,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "30d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2512.24867",
      "title": "Encyclo-K: Evaluating LLMs with Dynamically Composed Knowledge Statements",
      "abstract": "Benchmarks play a crucial role in tracking the rapid advancement of large language models (LLMs) and identifying their capability boundaries. However, existing benchmarks predominantly curate questions at the question level, suffering from three fundamental limitations: vulnerability to data contamination, restriction to single-knowledge-point assessment, and reliance on costly domain expert annotation. We propose Encyclo-K, a statement-based benchmark that rethinks benchmark construction from the ground up. Our key insight is that knowledge statements, not questions, can serve as the unit of curation, and questions can then be constructed from them. We extract standalone knowledge statements from authoritative textbooks and dynamically compose them into evaluation questions through random sampling at test time. This design directly addresses all three limitations: the combinatorial space is too vast to memorize, and model rankings remain stable across dynamically generated question sets, enabling reliable periodic dataset refresh; each question aggregates 8-10 statements for comprehensive multi-knowledge assessment; annotators only verify formatting compliance without requiring domain expertise, substantially reducing annotation costs. Experiments on over 50 LLMs demonstrate that Encyclo-K poses substantial challenges with strong discriminative power. Even the top-performing OpenAI-GPT-5.1 achieves only 62.07% accuracy, and model performance displays a clear gradient distribution--reasoning models span from 16.04% to 62.07%, while chat models range from 9.71% to 50.40%. These results validate the challenges introduced by dynamic evaluation and multi-statement comprehensive understanding. These findings establish Encyclo-K as a scalable framework for dynamic evaluation of LLMs' comprehensive understanding over multiple fine-grained disciplinary knowledge statements.",
      "authors": [
        "Yiming Liang",
        "Yizhi Li",
        "Yantao Du",
        "Ge Zhang",
        "Jiayi Zhou",
        "Yuchen Wu",
        "Yinzhu Piao",
        "Denghui Cao",
        "Tong Sun",
        "Ziniu Li"
      ],
      "published_at": "2025-12-31T13:55:54",
      "categories": [
        "cs.CL",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2512.24867v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.5302525162696838,
      "sim_memory": 0.38387632369995117,
      "sim_negative": 0.15658845007419586,
      "broad_relevance": 0.5302525162696838,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.6426062881946564,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.38387632369995117,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.48546669811010357,
      "memory_score": 0.21116289710998531,
      "quality_score": 0.595,
      "bridge_score": 0.21116289710998531,
      "max_axis_score": 0.48546669811010357,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2512.00902",
      "title": "Elastic Mixture of Rank-Wise Experts for Knowledge Reuse in Federated Fine-Tuning",
      "abstract": "Federated fine-tuning offers a promising solution for adapting Large Language Models (LLMs) to downstream tasks while safeguarding data privacy. However, its high computational and communication demands hinder its deployment on resource-constrained devices. In this paper, we propose SmartFed, a resource-efficient federated fine-tuning framework. SmartFed intelligently reuses knowledge embedded in existing LoRA modules, eliminating the need for expensive training from scratch when adapting LLMs to new tasks. To effectively exploit this knowledge and ensure scalability, we introduce the Mixture of Rank-Wise Experts (MoRE). MoRE decomposes LoRA modules into fine-grained rank-level experts. These experts are selectively activated and combined based on input semantics and resource budgets. Moreover, to optimize resource utilization, we present the Elastic Expert Quota Allocation (EEQA). EEQA adaptively allocates expert capacity across parameter matrices based on their contribution to model performance, focusing computing resources on the critical experts. Extensive evaluations across multiple benchmarks demonstrate that SmartFed significantly outperforms existing methods in model performance and training efficiency.",
      "authors": [
        "Yebo Wu",
        "Jingguang Li",
        "Zhijiang Guo",
        "Li Li"
      ],
      "published_at": "2025-11-30T14:09:59",
      "categories": [
        "cs.DC"
      ],
      "url": "http://arxiv.org/abs/2512.00902v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "systems",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.5359847545623779,
      "sim_memory": 0.44854536652565,
      "sim_negative": 0.29023200273513794,
      "broad_relevance": 0.5359847545623779,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.5890395045280457,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.44854536652565,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.7,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.47915135204792025,
      "memory_score": 0.305563609957695,
      "quality_score": 0.595,
      "bridge_score": 0.305563609957695,
      "max_axis_score": 0.47915135204792025,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2512.00290",
      "title": "EduEval: A Hierarchical Cognitive Benchmark for Evaluating Large Language Models in Chinese Education",
      "abstract": "Large language models (LLMs) demonstrate significant potential for educational applications. However, their unscrutinized deployment poses risks to educational standards, underscoring the need for rigorous evaluation. We introduce EduEval, a comprehensive hierarchical benchmark for evaluating LLMs in Chinese K-12 education. This benchmark makes three key contributions: (1) Cognitive Framework: We propose the EduAbility Taxonomy, which unifies Bloom's Taxonomy and Webb's Depth of Knowledge to organize tasks across six cognitive dimensions including Memorization, Understanding, Application, Reasoning, Creativity, and Ethics. (2) Authenticity: Our benchmark integrates real exam questions, classroom conversation, student essays, and expert-designed prompts to reflect genuine educational challenges; (3) Scale: EduEval comprises 24 distinct task types with over 11,000 questions spanning primary to high school levels. We evaluate 14 leading LLMs under both zero-shot and few-shot settings, revealing that while models perform well on factual tasks, they struggle with classroom dialogue classification and exhibit inconsistent results in creative content generation. Interestingly, several open source models outperform proprietary systems on complex educational reasoning. Few-shot prompting shows varying effectiveness across cognitive dimensions, suggesting that different educational objectives require tailored approaches. These findings provide targeted benchmarking metrics for developing LLMs specifically optimized for diverse Chinese educational tasks.",
      "authors": [
        "Guoqing Ma",
        "Jia Zhu",
        "Hanghui Guo",
        "Weijie Shi",
        "Yue Cui",
        "Jiawei Shen",
        "Zilong Li",
        "Yidan Liang"
      ],
      "published_at": "2025-11-29T03:09:50",
      "categories": [
        "cs.CL",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2512.00290v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.4835221767425537,
      "sim_memory": 0.31749534606933594,
      "sim_negative": 0.14464423060417175,
      "broad_relevance": 0.4835221767425537,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.6792082786560059,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.31749534606933594,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.47693789482116694,
      "memory_score": 0.2182486038208008,
      "quality_score": 0.595,
      "bridge_score": 0.2182486038208008,
      "max_axis_score": 0.47693789482116694,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2603.27490",
      "title": "AgentSwing: Adaptive Parallel Context Management Routing for Long-Horizon Web Agents",
      "abstract": "As large language models (LLMs) evolve into autonomous agents for long-horizon information-seeking, managing finite context capacity has become a critical bottleneck. Existing context management methods typically commit to a single fixed strategy throughout the entire trajectory. Such static designs may work well in some states, but they cannot adapt as the usefulness and reliability of the accumulated context evolve during long-horizon search. To formalize this challenge, we introduce a probabilistic framework that characterizes long-horizon success through two complementary dimensions: search efficiency and terminal precision. Building on this perspective, we propose AgentSwing, a state-aware adaptive parallel context management routing framework. At each trigger point, AgentSwing expands multiple context-managed branches in parallel and uses lookahead routing to select the most promising continuation. Experiments across diverse benchmarks and agent backbones show that AgentSwing consistently outperforms strong static context management methods, often matching or exceeding their performance with up to $3\\times$ fewer interaction turns while also improving the ultimate performance ceiling of long-horizon web agents. Beyond the empirical gains, the proposed probabilistic framework provides a principled lens for analyzing and designing future context management strategies for long-horizon agents.",
      "authors": [
        "Zhaopeng Feng",
        "Liangcai Su",
        "Zhen Zhang",
        "Xinyu Wang",
        "Xiaotian Zhang",
        "Xiaobin Wang",
        "Runnan Fang",
        "Qi Zhang",
        "Baixuan Li",
        "Shihao Cai"
      ],
      "published_at": "2026-03-29T02:59:27",
      "categories": [
        "cs.CL",
        "cs.AI",
        "cs.MA"
      ],
      "url": "http://arxiv.org/abs/2603.27490v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.5417897701263428,
      "sim_memory": 0.37209612131118774,
      "sim_negative": 0.2583639621734619,
      "broad_relevance": 0.5417897701263428,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.5866723358631134,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.37209612131118774,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.25,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.48053778141736986,
      "memory_score": 0.23262883639335633,
      "quality_score": 0.595,
      "bridge_score": 0.23262883639335633,
      "max_axis_score": 0.48053778141736986,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "30d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2512.01786",
      "title": "Who Judges the Judge? LLM Jury-on-Demand: Building Trustworthy LLM Evaluation Systems",
      "abstract": "As Large Language Models (LLMs) become integrated into high-stakes domains, there is a growing need for evaluation methods that are both scalable for real-time deployment and reliable for critical decision-making. While human evaluation is reliable, it is slow and costly. Single LLM judges are biased, and static juries lack adaptability. To overcome these limitations, we propose LLM Jury-on-Demand - a dynamic, learning-based framework for scalable and context-aware evaluation. Our method trains a set of reliability predictors to assess when LLM judges will agree with human experts, leveraging token distributions, embeddings, and structural input features. This enables a fully adaptive evaluation where, for each data point, an optimal jury of the most reliable judges is dynamically selected, and their scores are aggregated using their reliability as weights. Experiments on summarization and RAG benchmarks show that our dynamic jury system achieves significantly higher correlation with human judgment than both single-judge and static-jury baselines. These results highlight the promise of adaptive, learning-based juries for building scalable, more reliable and trustworthy evaluation systems for modern LLMs in high-stakes domains.",
      "authors": [
        "Xiaochuan Li",
        "Ke Wang",
        "Girija Gouda",
        "Shubham Choudhary",
        "Yaqun Wang",
        "Linwei Hu",
        "Joel Vaughan",
        "Freddy Lecue"
      ],
      "published_at": "2025-12-01T15:26:20",
      "categories": [
        "cs.AI",
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2512.01786v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.49739870429039,
      "sim_memory": 0.3284987807273865,
      "sim_negative": 0.22144091129302979,
      "broad_relevance": 0.49739870429039,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.6387252807617188,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.3284987807273865,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.4750284034013748,
      "memory_score": 0.22154963421821594,
      "quality_score": 0.595,
      "bridge_score": 0.22154963421821594,
      "max_axis_score": 0.4750284034013748,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2512.01983",
      "title": "Feature-Based Semantics-Aware Scheduling for Energy-Harvesting Federated Learning",
      "abstract": "Federated Learning (FL) on resource-constrained edge devices faces a critical challenge: The computational energy required for training Deep Neural Networks (DNNs) often dominates communication costs. However, most existing Energy-Harvesting FL (EHFL) strategies fail to account for this reality, resulting in wasted energy due to redundant local computations. For efficient and proactive resource management, algorithms that predict local update contributions must be devised. We propose a lightweight client scheduling framework using the Version Age of Information (VAoI), a semantics-aware metric that quantifies update timeliness and significance. Crucially, we overcome VAoI's typical prohibitive computational cost, which requires statistical distance over the entire parameter space, by introducing a feature-based proxy. This proxy estimates model redundancy using intermediate-layer extraction from a single forward pass, dramatically reducing computational complexity. Experiments conducted under extreme non-IID data distributions and scarce energy availability demonstrate superior learning performance while achieving energy reduction compared to existing baseline selection policies. Our framework establishes semantics-aware scheduling as a practical and vital solution for EHFL in realistic scenarios where training costs dominate transmission costs.",
      "authors": [
        "Eunjeong Jeong",
        "Giovanni Perin",
        "Howard H. Yang",
        "Nikolaos Pappas"
      ],
      "published_at": "2025-12-01T18:40:26",
      "categories": [
        "cs.LG",
        "cs.DC",
        "cs.IT",
        "cs.NI",
        "eess.SP"
      ],
      "url": "http://arxiv.org/abs/2512.01983v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.5101743936538696,
      "sim_memory": 0.4364962577819824,
      "sim_negative": 0.3040136396884918,
      "broad_relevance": 0.5101743936538696,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.5960915386676788,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.4364962577819824,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.2,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.47246604889631266,
      "memory_score": 0.28394887733459473,
      "quality_score": 0.595,
      "bridge_score": 0.28394887733459473,
      "max_axis_score": 0.47246604889631266,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2512.24601",
      "title": "Recursive Language Models",
      "abstract": "We study allowing large language models (LLMs) to process arbitrarily long prompts through the lens of inference-time scaling. We propose Recursive Language Models (RLMs), a general inference paradigm that treats long prompts as part of an external environment and allows the LLM to programmatically examine, decompose, and recursively call itself over snippets of the prompt. We find that RLMs can successfully process inputs up to two orders of magnitude beyond model context windows and, even for shorter prompts, dramatically outperform the quality of vanilla frontier LLMs and common long-context scaffolds across four diverse long-context tasks while having comparable cost. At a small scale, we post-train the first natively recursive language model. Our model, RLM-Qwen3-8B, outperforms the underlying Qwen3-8B model by $28.3\\%$ on average and even approaches the quality of vanilla GPT-5 on three long-context tasks. Code is available at https://github.com/alexzhang13/rlm.",
      "authors": [
        "Alex L. Zhang",
        "Tim Kraska",
        "Omar Khattab"
      ],
      "published_at": "2025-12-31T03:43:41",
      "categories": [
        "cs.AI",
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2512.24601v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.6256693005561829,
      "sim_memory": 0.49101313948631287,
      "sim_negative": 0.23632369935512543,
      "broad_relevance": 0.6256693005561829,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.46118026971817017,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.49101313948631287,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.2,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.4868778306245804,
      "memory_score": 0.2733039418458938,
      "quality_score": 0.595,
      "bridge_score": 0.2733039418458938,
      "max_axis_score": 0.4868778306245804,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.22623",
      "title": "SYMPHONY: Synergistic Multi-agent Planning with Heterogeneous Language Model Assembly",
      "abstract": "Recent advancements have increasingly focused on leveraging large language models (LLMs) to construct autonomous agents for complex problem-solving tasks. However, existing approaches predominantly employ a single-agent framework to generate search branches and estimate rewards during Monte Carlo Tree Search (MCTS) planning. This single-agent paradigm inherently limits exploration capabilities, often resulting in insufficient diversity among generated branches and suboptimal planning performance. To overcome these limitations, we propose Synergistic Multi-agent Planning with Heterogeneous langauge model assembly (SYMPHONY), a novel multi-agent planning framework that integrates a pool of heterogeneous language model-based agents. By leveraging diverse reasoning patterns across agents, SYMPHONY enhances rollout diversity and facilitates more effective exploration. Empirical results across multiple benchmark tasks show that SYMPHONY achieves strong performance even when instantiated with open-source LLMs deployable on consumer-grade hardware. When enhanced with cloud-based LLMs accessible via API, SYMPHONY demonstrates further improvements, outperforming existing state-of-the-art baselines and underscoring the effectiveness of heterogeneous multi-agent coordination in planning tasks.",
      "authors": [
        "Wei Zhu",
        "Zhiwen Tang",
        "Kun Yue"
      ],
      "published_at": "2026-01-30T06:26:34",
      "categories": [
        "cs.AI",
        "cs.MA"
      ],
      "url": "http://arxiv.org/abs/2601.22623v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.592624306678772,
      "sim_memory": 0.3191952705383301,
      "sim_negative": 0.19745001196861267,
      "broad_relevance": 0.592624306678772,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.5218358933925629,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.3191952705383301,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.486062676012516,
      "memory_score": 0.19175858116149902,
      "quality_score": 0.595,
      "bridge_score": 0.19175858116149902,
      "max_axis_score": 0.486062676012516,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2511.00640",
      "title": "DTS: Enhancing Large Reasoning Models via Decoding Tree Sketching",
      "abstract": "Large Reasoning Models (LRMs) achieve remarkable inference-time improvements through parallel thinking. However, existing methods rely on redundant sampling of reasoning trajectories, failing to effectively explore the reasoning space to uncover high-quality solutions. To address these limitations, we propose Decoding Tree Sketching (DTS), a plug-and-play decoding framework for structural multi-trajectory exploration and reasoning selection. For reasoning exploration, DTS sketches a backbone tree of the reasoning space by selectively branching at decision tokens. For reasoning selection, guided by length-accuracy anti-correlation, DTS designs an early termination to prioritize short and reliable trajectories during decoding. Experimental results across four LRMs and datasets demonstrate that DTS significantly enhances accuracy by 14% and reduces repetitive generation by 8% on average. Notably, DTS enables smaller models to outperform larger models with 10$\\times$ the size, highlighting its potential to strengthen reasoning capabilities.",
      "authors": [
        "Zicheng Xu",
        "Xiuyi Lou",
        "Guanchu Wang",
        "Yu-Neng Chuang",
        "Feng Luo",
        "Guangyao Zheng",
        "Alexander S. Szalay",
        "Zirui Liu",
        "Vladimir Braverman"
      ],
      "published_at": "2025-11-01T17:41:28",
      "categories": [
        "cs.AI",
        "cs.CL",
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2511.00640v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.5694093704223633,
      "sim_memory": 0.32168447971343994,
      "sim_negative": 0.2213895320892334,
      "broad_relevance": 0.5694093704223633,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.5490879118442535,
      "evidence_score": 0.4,
      "direct_memory_relevance": 0.32168447971343994,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7999999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.483185997903347,
      "memory_score": 0.19250534391403196,
      "quality_score": 0.595,
      "bridge_score": 0.19250534391403196,
      "max_axis_score": 0.483185997903347,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [],
      "status": "Monitor",
      "why_now": "",
      "why_not_higher": "",
      "downgrade_reasons": [],
      "what_would_raise_priority": "",
      "one_sentence_episode_hook": "",
      "source": "digest",
      "added": "",
      "issue_number": null
    }
  ],
  "deferred": [],
  "out_of_scope": [
    {
      "arxiv_id": "2602.23828",
      "title": "GenDRAM:Hardware-Software Co-Design of General Platform in DRAM",
      "abstract": "Dynamic programming (DP) algorithms, such as All-Pairs Shortest Path (APSP) and genomic sequence alignment, are fundamental to many scientific domains but are severely bottlenecked by data movement on conventional architectures. While Processing-in-Memory (PIM) offers a promising solution, existing accelerators often address only a fraction of the work-flow, creating new system-level bottlenecks in host-accelerator communication and off-chip data streaming. In this work, we propose GenDRAM, a massively parallel PIM accelerator that overcomes these limitations. GenDRAM leverages the immense capacity and internal bandwidth of monolithic 3D DRAM(M3D DRAM) to integrate entire data-intensive pipelines, such as the full genomics workflow from seeding to alignment, onto a single heterogeneous chip. At its core is a novel architecture featuring specialized Search PUs for memory-intensive tasks and universal, multiplier-less Compute PUs for diverse DP calculations. This is enabled by a 3D-aware data mapping strategy that exploits the tiered latency of M3D DRAM for performance optimization. Through comprehensive simulation, we demonstrate that GenDRAM achieves a transformative performance leap, outperforming state-of-the-art GPU systems by over 68x on APSP and over 22x on the end-to-end genomics pipeline.",
      "authors": [
        "Tsung-Han Lu",
        "Weihong Xu",
        "Tajana Rosing"
      ],
      "published_at": "2026-02-27T09:09:15",
      "categories": [
        "cs.AR"
      ],
      "url": "http://arxiv.org/abs/2602.23828v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "hardware",
      "domain_bucket": "bio",
      "paper_type": "empirical",
      "narrow_domain_flag": true,
      "sim_public": 0.40678226947784424,
      "sim_memory": 0.4826856255531311,
      "sim_negative": 0.21437117457389832,
      "broad_relevance": 0.40678226947784424,
      "momentum": 0.0,
      "teachability": 0.7,
      "novelty_score": 0.5705396234989166,
      "evidence_score": 0.34,
      "direct_memory_relevance": 0.4826856255531311,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.5,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.6000000000000001,
      "transferability_score": 0.3,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.22761562436819072,
      "memory_score": 0.5098056876659394,
      "quality_score": 0.52,
      "bridge_score": 0.22761562436819072,
      "max_axis_score": 0.5098056876659394,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Memory/Storage Adjacent",
        "Systems",
        "Hardware",
        "Application"
      ],
      "status": "Out of scope",
      "why_now": "It speaks to a real systems pressure point: data movement is dominating performance and energy, and near-memory design is getting renewed attention. The hardware-software co-design angle is timely for anyone watching bandwidth and locality limits.",
      "why_not_higher": "This is not really about modern AI training or inference, and the workloads are dynamic programming and genomics rather than large-model serving. The claimed gains come from simulation on a specialized platform, which limits immediate relevance and transfer.",
      "downgrade_reasons": [
        "Not centered on modern AI systems",
        "Memory link is architectural but not AI-memory specific",
        "Workloads are narrow and domain-specific",
        "Evidence is simulation-heavy on custom hardware"
      ],
      "what_would_raise_priority": "A stronger case would show how the architecture or techniques transfer to transformer-era AI bottlenecks such as KV-cache serving, model offload, embedding retrieval, or bandwidth-bound inference.",
      "one_sentence_episode_hook": "What if the real fix for data-hungry AI-era workloads is not a faster GPU, but moving the whole pipeline into DRAM itself?",
      "source": "digest",
      "added": "",
      "issue_number": null
    }
  ]
}