AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

D-ARL: A Distribution-Matched Asynchronous Reinforcement Learning Framework for Language Reasoning

Yinqi Bai, Tong Xialiang, Jie Wang, Hongyu Liu, Longdi Pan, Jiashuo Li, Zehao Wang, Jianye HAO, Mingxuan Yuan, Feng Wu

ICML 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning

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 distribution-matched sample selection (ablation)D-ARL (Qwen3-1.7B) · six reasoning benchmarks, averageaccuracy (%): 60.4 -> 65.2Improved
add multi-behavior policy optimization (ablation)D-ARL (Qwen3-1.7B) · six reasoning benchmarks, averageaccuracy (%): 55.4 -> 65.2Improved
add both selection and multi-behavior optimization (ablation)D-ARL (Qwen3-1.7B) · HumanEvalaccuracy (%): 63.6 -> 81.1Improved
multi-behavior optimization alone without selection (ablation)D-ARL (Qwen3-1.7B) · MATH-500accuracy (%): w/o Optimization 75.7 vs w/o both 72.2; w/o Selection 75.3Mixed
use D-ARL instead of CISPOQwen3-1.7B · six benchmarks, averageaccuracy (%): 58.8 -> 65.2Improved
use D-ARL instead of CISPOQwen3-4B · six benchmarks, averageaccuracy (%): 68.7 -> 74.3Improved
use D-ARL instead of DecoupledQwen3-1.7B · AIME24accuracy (%): 26.7 -> 26.7No clear effect
use D-ARL instead of synchronous GRPOQwen3-1.7B · six benchmarks, averageaccuracy (%): 53.0 -> 65.2Improved
naive asynchronous ARLHF with IS correction vs synchronous GRPOARLHF · six benchmarks, average (Qwen3-1.7B)accuracy (%): 53.0 -> 52.4Hurt
use D-ARL instead of SOTA asynchronous baseline for sample efficiencyQwen3-1.7B · six benchmarkstraining steps to target accuracy: 34.7% average improvementImproved
use D-ARL instead of SOTA baseline for wall-clock timeQwen3-1.7B · six benchmarkstraining time: 37.2% average reductionImproved
sample selection overheadD-ARL · GSM8K, LightEval, AIME24 (Qwen3-1.7B)share of training time (%): 13.7 average (Qwen3-4B: 13.3)No 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