AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

BandPO: Bridging Trust Regions and Ratio Clipping via Probability-Aware Bounds for LLM Reinforcement Learning

Yuan Li, Bo Wang, Yufei Gao, Yao Yuqian, Xinyuan Wang, Zhangyue Yin, Xipeng Qiu

ICML 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants · 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
replace fixed ratio clipping with Band probability-aware bounds (KL, delta 0.05)GRPO · AMC2023/AIME2024/AIME2025 average, Qwen2.5-3B-Instructmean@32: 16.82 -> 22.16Improved
replace canonical clipping with BandGRPO · average over three math benchmarks, 1.5B to 8B modelsmean@32: at least +1.5 points on every scaleImproved
use Band instead of Clip-HigherGRPO · average over three math benchmarksmean@32: +0.45, +2.11, +2.41, +0.80 across 1.5B, 3B, 7B, 8BImproved
use Band instead of DCPO-DACGRPO · average over three math benchmarksmean@32: +1.39, +0.80, +1.14, +0.57 across 1.5B, 3B, 7B, 8BImproved
use Band instead of canonical clippingGRPO · average pass@32, 1.5B, 3B, 8B modelspass@32: +5.13, +10.38, +3.48Improved
use Band instead of canonical clippingGRPO · average pass@32, DeepSeek-R1-Distill-Qwen-7Bpass@32: 66.76 -> 66.05No clear effect
heuristically relax the Band bound to cover the Clip-Higher range (ablation)BandPO (KL, 0.05) · average over three math benchmarksmean@32: original Band better by +1.07, +1.91, +0.96, +0.61 (1.5B, 3B, 7B, 8B)Hurt
sweep trust-region radius delta among 0.03, 0.05, 0.10 (ablation)BandPO, Qwen2.5-3B-Instruct · average over three math benchmarksmean@32 / pass@32: 0.05 gives 22.16/42.32 vs 20.28 (0.03) and 20.35 (0.10) mean@32Improved
sweep trust-region radius delta among 0.03, 0.05, 0.10 (ablation)BandPO, DeepSeek-R1-Distill-Qwen-7B · average over three math benchmarksmean@32: 49.78 / 50.71 / 50.28 for 0.03 / 0.05 / 0.10No clear effect
use Band bounds instead of fixed clipping for low-probability tokensGRPO, Qwen2.5-3B · clip-high rate for tokens with p < 0.2 and positive advantageclip rate: 20-60% -> approximately 0%Improved
use Band bounds instead of fixed clippingGRPO, Qwen2.5-3B · policy entropy during trainingentropyImproved
use Band instead of fixed clippingGRPO, Qwen2.5-3B-Instruct · end-to-end training timetotal hours: 39.67 -> 37.56 hoursImproved

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