AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

The Easy, the Hard, and the Learnable: Confidence and Difficulty-Adaptive Policy Optimization for LLM Reasoning

Zhanke Zhou, Xiangyu Lu, Chentao Cao, Brando Miranda, Tongliang Liu, Bo Han, Sanmi Koyejo

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
add CoDaPO (CoDaWeighting, CoDaSampling, CoDaLearning) to GRPOGRPO, Qwen2.5-Math-1.5B · 7 math benchmarks averageaverage accuracy (%): 39.08 -> 41.30Improved
add CoDaPO to GRPO on Qwen2.5-Math-7BGRPO, Qwen2.5-Math-7B · 7 math benchmarks averageaverage accuracy (%): 44.58 -> 46.67Improved
add CoDaPO to GRPO on Llama-3.2-1B-InstructGRPO, Llama-3.2-1B-Instruct · 7 math benchmarks averageaverage accuracy (%): 13.48 -> 15.05Improved
add CoDaWeighting to GRPO (ablation)GRPO, Qwen2.5-Math-1.5B · MATH500/AIME24/AIME25 averageaverage accuracy (%): 30.44 -> 31.53Improved
add CoDaSampling on top of CoDaWeighting (ablation)CoDaPO, Qwen2.5-Math-1.5B · MATH500/AIME24/AIME25 averageaverage accuracy (%): 31.53 -> 32.79Improved
add confidence term Vc(c)=c to difficulty-only weighting (ablation)CoDaWeighting · MATH500/AIME24/AIME25 averageaverage accuracy (%): 31.32 -> 31.53Improved
add difficulty term Vd to confidence-only weighting (ablation)CoDaWeighting · MATH500/AIME24/AIME25 averageaverage accuracy (%): 30.99 -> 31.53Improved
use Top-K=4 instead of Top-K=1 in CoDaSampling (ablation)CoDaPO · MATH500/AIME24/AIME25 averageaverage accuracy (%): 30.40 -> 32.79Improved
vary Top-K among 2, 4, 8 (ablation)CoDaSampling · MATH500/AIME24/AIME25 averageaverage accuracy (%): 32.48 / 32.79 / 32.62No clear effect
apply CoDaLearning to DAPODAPO · 7 math benchmarks averageaverage accuracy (%): 39.49 -> 41.03Improved
apply CoDaLearning to GPGGPG · 7 math benchmarks averageaverage accuracy (%): 39.77 -> 39.95Improved
use easy-to-hard curriculum instead of uniform samplingGRPO · 7 math benchmarks averageaverage accuracy (%): 39.08 -> 39.30No 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