AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Replay Failures as Successes: Sample-Efficient Reinforcement Learning for Instruction Following

Kongcheng Zhang, QI YAO, Shunyu Liu, Wenjian Zhang, Min Cen, Yang Zhou, Wenkai Fang, Yiru Zhao, Baisheng Lai, Mingli Song

ICML 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning · 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
add adaptive select-then-rewrite replay selection instead of random replay (ablation)HiR Llama-3.2-3B-Instruct · IFEvalinstruction-level accuracy: 79.9 -> 83.6Improved
add adaptive select-then-rewrite replay selection instead of random replay (ablation)HiR Qwen2.5-7B-Instruct · IFEvalinstruction-level accuracy: 79.5 -> 81.0Improved
add adaptive select-then-rewrite replay selection instead of random replay (ablation)HiR Qwen3-4B-Instruct-2507 · IFEvalinstruction-level accuracy: 85.2 -> 86.3Improved
add adaptive replay selection instead of random replay (ablation)HiR Qwen2.5-7B-Instruct · FollowBenchinstruction-level accuracy: 66.2 -> 65.1Hurt
add adaptive replay selection instead of random replay (ablation)HiR Llama-3.2-3B-Instruct · IFBenchinstruction-level accuracy: 28.2 -> 30.4Improved
add hindsight instruction replay to RL with constraint-level rewardRL-CR Llama-3.2-3B-Instruct · IFEvalinstruction-level accuracy: 79.1 -> 83.6Improved
add hindsight instruction replay to RL with constraint-level rewardRL-CR Qwen3-4B-Instruct-2507 · IFBenchinstruction-level accuracy: 36.9 -> 40.5Improved
train with DPO instead of base modelQwen2.5-7B-Instruct · IFEvalinstruction-level accuracy: 72.6 -> 66.9Hurt
train with SFT instead of base modelQwen2.5-7B-Instruct · CFBenchinstruction-level accuracy: 57.5 -> 53.1Hurt
train HiR on instruction following dataQwen3-4B-Instruct-2507 · MMLU-Pro (out-of-domain)accuracy: 69.6 -> 67.2Hurt
train HiR on instruction following dataQwen3-4B-Instruct-2507 · MATH-500 (out-of-domain)accuracy: 86.8 -> 88.2No clear effect
train HiR on instruction following dataQwen2.5-7B-Instruct · MATH-500 (out-of-domain)accuracy: 76.6 -> 76.6No clear effect

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