Add daily research log: SLOT, N-gram Residual Learning, LieQ (2026-03-25)

[sunnypatneedi]

Summary

• Adds logs/daily_research.md — 2026-03-25 deep arXiv scan covering three new applicable techniques:
  • SLOT (arXiv:2505.12392): Sample-specific δ vector on final hidden layer, few-step optimization per sample — legal TTT variant complementary to n-gram cache
  • N-gram Residual Learning (arXiv:2210.14431): Train neural model on residual distribution (what n-gram can't predict) — could tighten neural+n-gram combination by 0.01-0.03 bpb
  • LieQ (arXiv:2508.03332): Layer-wise mixed-precision PTQ (int7 critical / int5 redundant) — could recover 0.002-0.005 bpb or free ~200KB for larger n-gram cache
• Adds logs/experiments.md — full run history from sessions 1-3 with hypotheses and verdicts
• Updates .gitignore to track logs/daily_research.md and logs/experiments.md while keeping binary model artifacts excluded

Context

Research was performed against the competition leaderboard state as of 2026-03-25. Key context: n-gram eval cache confirmed legal, two-pass rescoring confirmed illegal. Open PR frontier at 0.5466 bpb (PR openai#798, order-adaptive entropy gating).

🤖 Generated with Claude Code