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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add distribution-matched sample selection (ablation) | D-ARL (Qwen3-1.7B) · six reasoning benchmarks, average | accuracy (%): 60.4 -> 65.2 | Improved |
| add multi-behavior policy optimization (ablation) | D-ARL (Qwen3-1.7B) · six reasoning benchmarks, average | accuracy (%): 55.4 -> 65.2 | Improved |
| add both selection and multi-behavior optimization (ablation) | D-ARL (Qwen3-1.7B) · HumanEval | accuracy (%): 63.6 -> 81.1 | Improved |
| multi-behavior optimization alone without selection (ablation) | D-ARL (Qwen3-1.7B) · MATH-500 | accuracy (%): w/o Optimization 75.7 vs w/o both 72.2; w/o Selection 75.3 | Mixed |
| use D-ARL instead of CISPO | Qwen3-1.7B · six benchmarks, average | accuracy (%): 58.8 -> 65.2 | Improved |
| use D-ARL instead of CISPO | Qwen3-4B · six benchmarks, average | accuracy (%): 68.7 -> 74.3 | Improved |
| use D-ARL instead of Decoupled | Qwen3-1.7B · AIME24 | accuracy (%): 26.7 -> 26.7 | No clear effect |
| use D-ARL instead of synchronous GRPO | Qwen3-1.7B · six benchmarks, average | accuracy (%): 53.0 -> 65.2 | Improved |
| naive asynchronous ARLHF with IS correction vs synchronous GRPO | ARLHF · six benchmarks, average (Qwen3-1.7B) | accuracy (%): 53.0 -> 52.4 | Hurt |
| use D-ARL instead of SOTA asynchronous baseline for sample efficiency | Qwen3-1.7B · six benchmarks | training steps to target accuracy: 34.7% average improvement | Improved |
| use D-ARL instead of SOTA baseline for wall-clock time | Qwen3-1.7B · six benchmarks | training time: 37.2% average reduction | Improved |
| sample selection overhead | D-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
- AREAL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning (NeurIPS 2025 · similarity 0.93)
- AsyncOPD: How Stale Can On-Policy Distillation Be? (NeurIPS 2026 · similarity 0.91)
- Asynchronous RLHF: Faster and More Efficient Off-Policy RL for Language Models (ICLR 2025 · similarity 0.91)
- Trajectory Balance with Asynchrony: Decoupling Exploration and Learning for Fast, Scalable LLM Post-Training (NeurIPS 2025 · similarity 0.90)
- Resource-Efficient Reinforcement for Reasoning Large Language Models via Dynamic One-Shot Policy Refinement (ICML 2026 · similarity 0.90)
- d1: Scaling Reasoning in Diffusion Large Language Models via Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
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.
