Research map / Reinforcement learning / RL from feedback
Stabilizing Reinforcement Learning for Diffusion Language Models
Jianyuan Zhong, Kaibo Wang, Ding Ding, Zijin Feng, Haoli Bai, Yang Xiang, Jiacheng Sun, Qiang Xu
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 unconditional clipping (vs self-normalization only) (ablation) | StableDRL · Countdown/GSM8K training reward | training stability | Improved |
| add self-normalization (vs clipping only) (ablation) | StableDRL · Countdown/GSM8K training reward | training stability | Improved |
| clipping threshold epsilon 5 instead of 1 (ablation) | StableDRL · GSM8K training reward | convergence and final reward | Improved |
| clipping threshold epsilon 100 or 1000 instead of 5 (ablation) | StableDRL · GSM8K training reward | training stability | Hurt |
| replace SPG with StableDRL at length 256 | LLaDA-8B-Instruct · Countdown | pass@1 accuracy: 70.7 -> 84.4 | Improved |
| replace SPG with StableDRL at length 256 | LLaDA-8B-Instruct · MATH500 | pass@1 accuracy: 40.0 -> 45.2 | Improved |
| replace ESPO with StableDRL | LLaDA-8B-Instruct · GSM8K avg | pass@1 accuracy: 82.0 -> 84.2 | Improved |
| replace ESPO with StableDRL | LLaDA-8B-Instruct · Sudoku avg | pass@1 accuracy: 86.0 -> 91.5 | Improved |
| replace SPG with StableDRL | LLaDA-8B-Instruct · Sudoku avg | pass@1 accuracy: 90.0 -> 91.5 | Improved |
| replace base SDAR-8B with StableDRL RL (static) | SDAR-8B · AIME 24 | pass@1 accuracy: 11.8 -> 16.7 | Improved |
| replace Trado with StableDRL (static) | SDAR-8B · AIME 24 | pass@1 accuracy: 13.3 -> 16.7 | Improved |
| replace Trado with StableDRL (dynamic) | SDAR-8B · AIME 24 | pass@1 accuracy: 11.0 -> 13.3 | Improved |
Closest papers by meaning
- d1: Scaling Reasoning in Diffusion Large Language Models via Reinforcement Learning (NeurIPS 2025 · similarity 0.91)
- Reinforcement Learning for Diffusion LLMs with Entropy-Guided Step Selection and Stepwise Advantages (NeurIPS 2026 · similarity 0.90)
- Stabilizing Policy Optimization via Logits Convexity (NeurIPS 2026 · similarity 0.90)
- On Predictability of Reinforcement Learning Dynamics for Large Language Models (ICLR 2026 · similarity 0.89)
- Block-R1: Rethinking the Role of Block Size in Multi-domain Reinforcement Learning for Diffusion Large Language Models (NeurIPS 2026 · similarity 0.89)
- DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Models (NeurIPS 2026 · similarity 0.89)
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.
