Research map / Large language models / Instruction and fine-tuning
AIR: Post-training Data Selection for Reasoning via Attention Head Influence
Jinrui Liu, Kai Hua, Xuanguang Pan, Ge Zhang, Yong Wang, Shuai Ma, Chongyang Tao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| select 1K samples with AIR instead of random | Qwen2.5-32B SFT, Gemini traces · s1K Gemini setting | average Pass@1: 47.85 -> 54.78 | Improved |
| select 1K samples with AIR instead of random | Qwen2.5-32B SFT, R1 traces · s1K-1.1 R1 setting | average Pass@1: 60.02 -> 67.14 | Improved |
| replace manual s1K-1.1 curation with AIR sample selection | Qwen2.5-32B SFT · s1K-1.1 R1 setting | average Pass@1: 65.43 -> 67.14 | Improved |
| replace manual s1K curation with AIR sample selection | Qwen2.5-32B SFT · s1K Gemini setting | average Pass@1: 56.48 -> 54.78 | Hurt |
| add AIR step-level loss weighting to s1K-1.1 training | s1K-1.1 SFT · s1K-1.1 R1 setting | average Pass@1: 65.43 -> 70.32 | Improved |
| select weighted steps by AIR instead of entropy (ablation) | step-weighted SFT · s1K-1.1 | 3-benchmark average Pass@1: 73.76 -> 75.98 | Improved |
| select weighted steps by AIR instead of randomly (ablation) | step-weighted SFT · s1K-1.1 | 3-benchmark average Pass@1: 70.76 -> 75.98 | Improved |
| weight random steps instead of no weighting (ablation) | s1K-1.1 SFT · s1K-1.1 | 3-benchmark average Pass@1: 70.57 -> 70.76 | No clear effect |
| weight high-entropy steps instead of no weighting (ablation) | s1K-1.1 SFT · s1K-1.1 | 3-benchmark average Pass@1: 70.57 -> 73.76 | Improved |
| use 20% critical steps rather than 40% or 50% (ablation) | step-level AIR SFT · s1K-1.1 | Pass@1: AIME 2024 66.67, GPQA 65.66, MATH500 95.60 at 20% | Improved |
| use weight multiplier 2x rather than 5x or 10x (ablation) | step-level AIR SFT · s1K-1.1 | Pass@1 | Improved |
| select with AIR instead of length heuristic (ablation) | Qwen2.5-32B SFT · s1K Gemini setting | 3-benchmark average Pass@1: 61.10 -> 65.27 | Improved |
Closest papers by meaning
- Stop When Further Reasoning Won’t Help: Attention-State Adaptive Generation in Reasoning Models (ICML 2026 · similarity 0.89)
- Selection, Reflection and Self-Refinement: Revisit Reasoning Tasks via a Causal Lens (ICLR 2026 · similarity 0.88)
- Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models (NeurIPS 2025 · similarity 0.88)
- Cognitive Mirrors: Exploring the Diverse Functional Roles of Attention Heads in LLM Reasoning (NeurIPS 2025 · similarity 0.88)
- Let's (not) just put things in Context: Test-time Training for Long-context LLMs (ICLR 2026 · similarity 0.87)
- DOTS: Learning to Reason Dynamically in LLMs via Optimal Reasoning Trajectories Search (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
