CURE-Med: Curriculum-Informed Reinforcement Learning for Multilingual Medical Reasoning

by

Eric Onyame* University of Virginia

Akash Ghosh* IIT-Patna

Subhadip Baidya IIT-Patna

Sriparna Saha IIT-Patna

Xiuying Chen MBZUAI

Chirag Agarwal University of Virginia

^*Equal contribution. Corresponding authors: Eric Onyame, Akash Ghosh

This repository hosts the codebase and dataset for CURE-Med, a framework for improving multilingual medical reasoning in large language models (LLMs). Below, we provide an overview of the project along with key training and implementation details.

Overview

Large language models (LLMs) perform strongly on monolingual math and commonsense reasoning, but they remain unreliable for multilingual medical reasoning—limiting safe use in real-world, multilingual healthcare settings. To address this, we introduce CUREMED-BENCH, a high-quality multilingual medical reasoning benchmark of open-ended questions with a single verifiable answer, spanning 13 languages, including under-represented languages such as Amharic, Yoruba, and Swahili. Building on this benchmark, we propose CURE-MED, a curriculum-informed reinforcement learning framework that combines code-switching-aware supervised fine-tuning with Group Relative Policy Optimization to improve both logical correctness and language stability. Across 13 languages, CURE-MED consistently outperforms strong baselines and scales effectively, reaching 85.21% language consistency and 54.35% logical correctness at 7B parameters, and 94.96% language consistency and 70.04% logical correctness at 32B parameters. Overall, our results move toward more reliable and equitable multilingual medical reasoning with LLMs.

Key Figure

Figure 1. CURE-MED pipeline: (A) clinically validated multilingual data curation (e.g., MedlinePlus), (B) code-switching-aware supervised fine-tuning of a Qwen2.5-Instruct backbone, and (C) GRPO-guided curriculum RL from high- to mid- to low-resource languages to improve logical correctness and language consistency.

_{High-resolution PDF: Figure 1}

For full technical details and experiments, see the paper on arXiv and the project website.

---

Dataset

• CUREMED-BENCH: Provided in data.zip. The dataset contains open-ended medical reasoning questions with a single verifiable answer across 13 languages.
  Unzip data.zip before running training or evaluation.

• Hugging Face: CUREMED-BENCH is also available on Hugging Face:
  https://huggingface.co/datasets/Aikyam-Lab/CUREMED-BENCH

---

Repository Structure

• baseline_inference/ — Baseline inference scripts for evaluation.
• SFT/ — Code-switching-aware supervised fine-tuning (SFT) training pipeline.
• SFT_Inference/ — Inference and evaluation for SFT checkpoints.
• Curriculum_RFT/ — Curriculum-informed reinforcement learning / RFT training (GRPO-guided).
• RFT_Inference/ — Inference and evaluation for RFT checkpoints.
• figures/ — Figures used in the README and paper.
• README.md — Project documentation.
• data.zip — Packaged dataset release for local use.
• datasets/ — Codeswitched dataset release for local use.

Supervised Fine-Tuning (SFT)

SFT code is in SFT/:

• code_switch_sft.py — code-switching-aware SFT training script
• deepspeed_zero3.yaml — DeepSpeed ZeRO-3 config
• code_switch_batch_script.sh — Slurm batch script for launching SFT

Environment

We ran SFT with Python 3.11.13. Ensure your environment includes: torch, transformers, datasets, trl, accelerate, deepspeed.

Data (SFT_data)

Provide SFT training files as JSONL under a directory like: /path/to/SFT_data/*.jsonl

Each example must contain: question, reasoning, answer, language.

Set the dataset path in the batch script via: --data_dir="/path/to/SFT_data".

Models & GPU Recommendations

We used Qwen2.5-Instruct variants: 1.5B, 3B, 7B, 14B, 32B.

• 1.5B / ~4B: recommended 4× A100
• 7B / 14B / 32B: recommended ≥ 8× A100

SFT Hyperparameters

• Optimizer: AdamW (β1=0.9, β2=0.999)
• LR: 1e-5 (cosine, warmup ratio 0.1), epochs: 3
• Effective batch size: 32, max seq length: 4096
• Precision: bf16, DeepSpeed ZeRO-3 + gradient checkpointing

Run (Slurm)

Edit SFT/code_switch_batch_script.sh:

• set base_model="Qwen/Qwen2.5-*-Instruct"
• set --data_dir="/path/to/SFT_data"
• request GPUs via #SBATCH --gres=gpu:a100:<N>
• match Accelerate processes to GPU count (e.g., --num_processes <N>)

Submit: sbatch SFT/code_switch_batch_script.sh

Reinforcement Fine-Tuning (RFT / GRPO Curriculum)

RFT is implemented in Curriculum_RFT/ as a 3-stage GRPO curriculum over staged datasets:

• Datasets: Curriculum_RFT/staged/{high,medium,low}/ (each contains *.jsonl)
• Training code: Curriculum_RFT/Training_Stages/{Stage_one_training,Stage_two_training,Stage_three_training}/

Environment

We ran RFT with Python 3.11.13 . All stages use full fine-tuning.

Dataset format

Each JSONL example must include: question, reasoning, answer, language.

Curriculum order (sequential checkpoints)

Run stages in this order (each stage initializes from the previous checkpoint):

1. Stage 1 (High-resource): start from the SFT checkpoint + staged/high/
2. Stage 2 (Medium-resource): start from Stage 1 checkpoint + staged/medium/
3. Stage 3 (Low-resource): start from Stage 2 checkpoint + staged/low/

Run RFT (Slurm)

Each stage provides a Slurm launcher script inside its stage folder. Update the script (or args) to point to the correct starting checkpoint and dataset path.

Note: Ensure accelerate --num_processes matches the number of GPUs requested in Slurm (e.g., 4 GPUs -> --num_processes 4). Output checkpoints are saved to the --output_dir specified in each stage script.

Inference: The SFT_Inference/ and RFT_Inference/ folders contain scripts for running inference with the trained SFT and RFT checkpoints. Please update the relevant model/checkpoint and data paths in the scripts before running.

Citation for CURE-Med Paper

Below is the BibTeX entry for the paper:

@misc{onyame2026curemedcurriculuminformedreinforcementlearning,
      title={CURE-Med: Curriculum-Informed Reinforcement Learning for Multilingual Medical Reasoning},
      author={Eric Onyame and Akash Ghosh and Subhadip Baidya and Sriparna Saha and Xiuying Chen and Chirag Agarwal},
      year={2026},
      eprint={2601.13262},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2601.13262},
}