AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

TAPO: Dynamic Teacher and Perturbed Answer Injection for Policy Optimization

Maowei Jiang, Zihang Wang, Qi Wang, Peter Búš, Moquan Cheng, Yifan Wang, Quangao Liu, Ruiqi Li, Pengyu Zeng, Ruikai Liu, Alan Liang, Yansong Xu and 3 more

AAAI 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants · 1 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 Dynamic Teacher Injection (ablation)GRPO · AIME 2024pass@1 accuracy (%): 34.3 -> 39.7Improved
add Perturbed Answer Injection after DTI (ablation)GRPO+DTI · AIME 2024pass@1 accuracy (%): 39.7 -> 42.8Improved
add InfoLen reward shaping (ablation)GRPO+DTI+PAI · AIME 2024pass@1 accuracy (%): 42.8 -> 44.3Improved
full TAPO versus naive GRPODeepSeek-R1-Distill-Qwen-1.5B · AIME 2024pass@1 accuracy (%): 34.3 -> 44.3Improved
full TAPO versus base modelDeepSeek-R1-Distill-Qwen-1.5B · AIME 2024pass@1 accuracy (%): 28.9 -> 44.3Improved
full TAPO versus naive GRPO, averageDeepSeek-R1-Distill-Qwen-1.5B · Six math benchmarks averageaverage accuracy (%): 56.86 -> 62.87Improved
full TAPO versus naive GRPODeepSeek-R1-Distill-Qwen-1.5B · Minerva Mathpass@1 accuracy (%): 28.2 -> 31.6Improved
InfoLen instead of length-only penalty (ablation)GRPO-style training · Output lengthmean tokens: 3187 -> 2793Improved
InfoLen versus length-only penalty accuracy (ablation)GRPO-style training · Math reasoningaccuracyImproved
advantage-level instead of reward-level intervention for injected samples (ablation)TAPO · Training reward curvestraining stabilityImproved
add DTI versus dynamic sampling (ablation)GRPO · Training rewardconvergence speedImproved
TAPO versus Qwen2.5-7B-SimpleRL1.5B model · Minerva Mathpass@1 accuracy (%): 39.7 vs 31.6Hurt

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