Research map / Reinforcement learning / RL from feedback
Improving Reasoning for Diffusion Language Models via Group Diffusion Policy Optimization
Kevin Rojas, Jiahe Lin, Kashif Rasul, Anderson Schneider, Yuriy Nevmyvaka, Molei Tao, Wei Deng
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 |
|---|---|---|---|
| replace diffu-GRPO token-level likelihood with GDPO sequence-level ELBO (SFT first) | LLaDA-8B-Instruct · Countdown (length 512) | accuracy: 42.2 -> 80.86 | Improved |
| replace diffu-GRPO with GDPO (SFT first) | LLaDA-8B-Instruct · GSM8K (length 512) | accuracy: 82.1 -> 84.99 | Improved |
| replace diffu-GRPO with GDPO (SFT first) | LLaDA-8B-Instruct · MATH500 (length 128) | accuracy: 33.8 -> 33.2 | No clear effect |
| replace diffu-GRPO with GDPO (SFT first) | LLaDA-8B-Instruct · Sudoku (length 512) | accuracy: 9.5 -> 26.17 | Improved |
| replace diffu-GRPO with GDPO | LLaDA-8B-Instruct · MBPP (length 256) | pass rate: 45.5 -> 50.6 | Improved |
| replace diffu-GRPO with GDPO | LLaDA-8B-Instruct · HumanEval (length 128) | pass rate: 29.3 -> 26.2 | Hurt |
| replace diffu-GRPO with GDPO | LLaDA-8B-Instruct · HumanEval (length 512) | pass rate: 34.8 -> 39.0 | Improved |
| RL fine-tuning with GDPO on coding without SFT | LLaDA-8B-Instruct · MBPP (length 256) | pass rate: 41.2 -> 50.6 | Improved |
| use 3 quadrature points instead of 2 (ablation) | GDPO · coding tasks | accuracy | Improved |
| use semi-deterministic quadrature ELBO instead of naive Monte Carlo (ablation) | GDPO · Countdown | test accuracy | Improved |
| use GDPO instead of diffu-GRPO at longer generation length | LLaDA-8B-Instruct · 512-token sequences | accuracy | Improved |
Closest papers by meaning
- Reinforcement Learning for Diffusion LLMs with Entropy-Guided Step Selection and Stepwise Advantages (NeurIPS 2026 · similarity 0.92)
- d2: Improving Reasoning in Diffusion Language Models via Trajectory Likelihood Estimation (ICML 2026 · similarity 0.90)
- MRO: Enhancing Reasoning in Diffusion Language Models via Multi-Reward Optimization (NeurIPS 2025 · similarity 0.90)
- d1: Scaling Reasoning in Diffusion Large Language Models via Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- Revolutionizing Reinforcement Learning Framework for Diffusion Large Language Models (ICLR 2026 · similarity 0.90)
- 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.
