AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

HiPO: Self-Hint Policy Optimization for RLVR

Deng Qiyuan, Kehai Chen, Min Zhang, Zhongwen Xu

ICLR 2026 · Reinforcement learning › RL from feedback · approach: Scaling RL with verifiable rewards

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 on-policy self-hint rolloutsGRPO · Average (5 benchmarks)avg@32: 49.8 -> 54.8Improved
add on-policy self-hint rolloutsGRPO · CMIMC 2025avg@32: 43.5 -> 53.8Improved
add on-policy self-hint rolloutsGRPO · BRUMO 2025avg@32: 41.7 -> 46.6Improved
add on-policy self-hint rolloutsGRPO · AIME 2024avg@32: 72.1 -> 76.7Improved
HiPO versus DAPO dynamic samplingDAPO · Average (5 benchmarks)avg@32: 53.7 -> 54.8Improved
add on-policy self-hint rolloutsGRPO · Average (5 benchmarks)maj@32: 57.0 -> 61.7Improved
use fixed hint ratio 0.05 instead of random ratio in [0.05,0.45] (ablation)HiPO · AIME 2024 / 2025 training curvesAvg@16Hurt
use fixed hint ratio 0.80 instead of random ratio (ablation)HiPO · AIME 2024 / 2025 training curvesAvg@16Hurt
use static off-policy hints from the base model (first 20% tokens) instead of on-policy hints (ablation)HiPO · AIME 2024 / 2025 training curvesAvg@16Hurt
add hint-based group replacementGRPO · Learnable group ratiolearnable groups: 17.5% -> 75.8%Improved
add on-policy self-hint rolloutsGRPO · HumanEval (OOD)Pass@1: 71.3 -> 71.3No clear effect
add on-policy self-hint rolloutsQwen3-8B · HumanEval (OOD)Pass@1: 70.7 -> 71.3Improved

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