Research map / Reinforcement learning / RL from feedback
SPG: Sandwiched Policy Gradient for Masked Diffusion Language Models
Chenyu Wang, Paria Rashidinejad, DiJia Su, Song Jiang, Sid Wang, Siyan Zhao, Cai Zhou, Shannon Zejiang Shen, Feiyu Chen, Tommi Jaakkola, Yuandong Tian, Bo Liu
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 the negative advantage loss (ablation) | SPG · GSM8K/MATH500/Countdown/Sudoku (avg lengths) | accuracy: 77.4 -> 83.1 (GSM8K); 32.7 -> 38.4 (MATH500); 45.5 -> 69.9 (Countdown); 68.8 -> 90.0 (Sudoku) with Mixture | Improved |
| use EUBO for negative advantage traces instead of ELBO (ablation) | SPG · GSM8K/Countdown/Sudoku (avg lengths) | accuracy: 80.9 -> 81.6; 67.1 -> 69.3; 82.4 -> 86.1 | Improved |
| use EUBO for negative advantage traces instead of ELBO (ablation) | SPG · MATH500 (avg lengths) | accuracy: 37.4 -> 36.7 | Hurt |
| use Mixture of ELBO and EUBO instead of ELBO for negative traces (ablation) | SPG · four reasoning benchmarks (avg lengths) | accuracy: +2.2 GSM8K, +1.0 MATH500, +2.8 Countdown, +7.6 Sudoku | Improved |
| use block-wise masking instead of random masking (ablation) | SPG w/ EUBO · Countdown (avg lengths) | accuracy: 45.4 -> 69.3 | Improved |
| use block-wise masking instead of random masking (ablation) | SPG w/ EUBO · MATH500 (avg lengths) | accuracy: 36.7 -> 36.7 | No clear effect |
| use block-wise masking instead of random masking (ablation) | SPG w/ Mixture · MATH500 and Countdown (avg lengths) | accuracy: 36.9 -> 38.4; 62.8 -> 69.9 | Improved |
| mix lower and upper bounds with omega in (0,1) (ablation) | SPG · benchmarks | accuracy | Improved |
| use intermediate beta between 0 and 1 (ablation) | SPG · benchmarks | accuracy | Hurt |
| replace UniGRPO with SPG w/ Mixture | dLLM RL (LLaDA-8B-Instruct) · GSM8K/MATH500/Countdown/Sudoku (len 256) | accuracy: 82.5 -> 86.1; 37.4 -> 40.0; 43.0 -> 70.7; 67.0 -> 94.0 | Improved |
| replace UniGRPO with SPG w/ Mixture | dLLM RL (LLaDA-8B-Instruct) · HumanEval (len 256) | accuracy: 39.6 -> 39.0 | Hurt |
| replace UniGRPO with SPG w/ Mixture | dLLM RL (LLaDA-8B-Instruct) · MBPP (len 256) | accuracy: 45.5 -> 50.6 | Improved |
Closest papers by meaning
- d2: Improving Reasoning in Diffusion Language Models via Trajectory Likelihood Estimation (ICML 2026 · similarity 0.90)
- Reinforcement Learning for Diffusion LLMs with Entropy-Guided Step Selection and Stepwise Advantages (NeurIPS 2026 · similarity 0.90)
- d1: Scaling Reasoning in Diffusion Large Language Models via Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- Revolutionizing Reinforcement Learning Framework for Diffusion Large Language Models (ICLR 2026 · similarity 0.88)
- DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Models (NeurIPS 2026 · similarity 0.87)
- Fine-Tuning Discrete Diffusion Models with Policy Gradient Methods (NeurIPS 2025 · similarity 0.87)
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.
