AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

DRPO: Efficient Reasoning via Decoupled Reward Policy Optimization

Gang Li, Yan Chen, Ming Lin, Tianbao Yang

ICLR 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants · 18 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 length reward to GRPO-style advantage (RLOO-LP, ALP, HAPO)GRPO-style methods · math average, 7Bpass@1 and AES: RLOO-LP: 38% length reduction with 7.1% performance lossHurt
decouple length reward by normalizing it only within correct rollouts (DRPO)DisCO · math average, 7Breasoning length and performance: 3053 -> 1502 tokens (51% reduction), 2.6% performance lossImproved
use DRPO instead of HAPODeepSeek-R1-Distill-Qwen-7B · math averagepass@1: 0.596 -> 0.714Improved
use DRPO instead of ALPDeepSeek-R1-Distill-Qwen-1.5B · math averageAES: -0.387 -> 0.178Improved
use DRPO instead of RLOO-LPDeepSeek-R1-Distill-Llama-8B · math averageAES: 0.251 -> 0.297Improved
apply DRPO on easy questionsDeepSeek-R1-Distill-Qwen-1.5B · GSM8Klength and accuracy: 1563 -> 356 tokens (77.2% reduction), -1.1% accuracyImproved
decrease lambda from 0.5 to 0.1 in DRPO (ablation)DRPO 7B · math averageAES: -0.007 -> 0.249Improved
decrease lambda from 0.5 to 0.1 in DRPO (ablation)DRPO 1.5B · AIMEAES: -0.143 -> -0.729Hurt
use concave or cosine length reward instead of linear in DRPO (ablation)DRPO 1.5B · math averageaccuracy vs length trade-offMixed
apply DRPO to non-math reasoningDeepSeek-R1-Distill-Qwen-1.5B · K&K logic puzzleslength and accuracy: 2095 -> 1400 tokens; accuracy 0.972 -> 0.974Improved

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