AI Paper Atlas

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

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning for reasoning data · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
select 1K samples with AIR instead of randomQwen2.5-32B SFT, Gemini traces · s1K Gemini settingaverage Pass@1: 47.85 -> 54.78Improved
select 1K samples with AIR instead of randomQwen2.5-32B SFT, R1 traces · s1K-1.1 R1 settingaverage Pass@1: 60.02 -> 67.14Improved
replace manual s1K-1.1 curation with AIR sample selectionQwen2.5-32B SFT · s1K-1.1 R1 settingaverage Pass@1: 65.43 -> 67.14Improved
replace manual s1K curation with AIR sample selectionQwen2.5-32B SFT · s1K Gemini settingaverage Pass@1: 56.48 -> 54.78Hurt
add AIR step-level loss weighting to s1K-1.1 trainings1K-1.1 SFT · s1K-1.1 R1 settingaverage Pass@1: 65.43 -> 70.32Improved
select weighted steps by AIR instead of entropy (ablation)step-weighted SFT · s1K-1.13-benchmark average Pass@1: 73.76 -> 75.98Improved
select weighted steps by AIR instead of randomly (ablation)step-weighted SFT · s1K-1.13-benchmark average Pass@1: 70.76 -> 75.98Improved
weight random steps instead of no weighting (ablation)s1K-1.1 SFT · s1K-1.13-benchmark average Pass@1: 70.57 -> 70.76No clear effect
weight high-entropy steps instead of no weighting (ablation)s1K-1.1 SFT · s1K-1.13-benchmark average Pass@1: 70.57 -> 73.76Improved
use 20% critical steps rather than 40% or 50% (ablation)step-level AIR SFT · s1K-1.1Pass@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-1.1Pass@1Improved
select with AIR instead of length heuristic (ablation)Qwen2.5-32B SFT · s1K Gemini setting3-benchmark average Pass@1: 61.10 -> 65.27Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy