AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Buffer Matters: Unleashing the Power of Off-Policy Reinforcement Learning in Large Language Model Reasoning

Xu Wan, Yansheng Wang, Wenqi Huang, Mingyang Sun

ICLR 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants

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 re-evaluated difficult samples X2 (ablation)BAPO · Countdown avg (Qwen2.5-Math-1.5B)accuracy: 47.81 -> 60.25Improved
add reused high-quality samples X3 (ablation)BAPO · Countdown avg (Qwen2.5-Math-1.5B)accuracy: 51.59 -> 60.25Improved
add re-evaluated difficult samples X2 (ablation)BAPO · Geometry3K avg (Qwen2.5-VL-3B)accuracy: 33.75 -> 43.22Improved
add reused high-quality samples X3 (ablation)BAPO · Geometry3K avg (Qwen2.5-VL-3B)accuracy: 36.01 -> 43.22Improved
replace GRPO with BAPO (off-policy batch adaptation)GRPO · Math avg (DeepSeek-R1-Distill-1.5B)accuracy: 51.58 -> 56.01Improved
use BAPO instead of DAPODAPO · Math avg (DeepSeek-R1-Distill-1.5B)accuracy and rollouts: 54.20 -> 56.01 accuracy; 1921k -> 733k rolloutsImproved
use BAPO instead of DAPO at 7BDAPO · Countdown avg (Qwen2.5-Math-7B)accuracy: 68.09 -> 68.13No clear effect
use BAPO instead of DAPO on MATH MinervaDAPO · Minerva (1.5B)accuracy: 30.70 -> 29.55Hurt
use naive off-policy methods (RePO, Remix-GRPO, delayed GRPO v=5) instead of GRPOGRPO · Math avg (DeepSeek-R1-Distill-1.5B)accuracy: 51.58 vs 51.57 / 50.54 / 50.53No clear effect
use fixed difficulty thresholds instead of adaptive boundaries (ablation)BAPO · Mini-test and fixed rangesaccuracyHurt
use rollout delay v and re-rollout frequency m (v=5, m=5 vs extreme delays) (ablation)BAPO · hyperparameter sweepaccuracyNo clear effect

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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