Research map / Image generation / Fast diffusion sampling
Improving Sampling for Masked Diffusion Models via Information Gain
Kaisen Yang, Jayden Teoh, Kaicheng Yang, Yitong Zhang, Alex Lamb
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 greedy entropy selection with Info-Gain Sampler | Dream-7B-Instruct · Reasoning avg, K=1 | average accuracy: 54.4 -> 62.0 | Improved |
| replace greedy entropy selection with Info-Gain Sampler | Dream-7B-Instruct · Reasoning avg, K=2 | average accuracy: 38.6 -> 53.4 | Improved |
| add information-gain look-ahead vs PC-Sampler | Dream-7B-Instruct · Reasoning avg, K=2 | average accuracy: 49.8 -> 53.4 | Improved |
| add information-gain look-ahead | SDAR-8B-Chat · Math+code avg, K=1 | average accuracy: 54.6 -> 66.2 | Improved |
| add information-gain look-ahead | TraDo-8B-Instruct · Math+code avg, K=1 | average accuracy: 56.9 -> 68.2 | Improved |
| use Info-Gain instead of LookUM | Dream-7B-Instruct · MATH500, K=2 | accuracy: 35.2 -> 34.2 | Hurt |
| use Info-Gain instead of LookUM | Dream-7B-Instruct · MATH500, K=1 | accuracy: 51.4 -> 51.3 | No clear effect |
| use Info-Gain instead of Margin sampling | MMaDa · GenEval avg | score: 56.3 -> 58.2 | Improved |
| use Info-Gain instead of Confidence sampling | MMaDa · ImageNet-512 | FID: 43.3 -> 38.1 | Improved |
| use Info-Gain at high token temperature 1.5, K=2 | SDAR-8B-Chat · AlpacaEval | win rate vs Entropy: 80.3 | Improved |
| let the sampler resolve reasoning before answer (AR order vs greedy entropy) | MDM greedy samplers · Binary judgment case study | accuracy: 67 -> 90 | Improved |
| use Info-Gain beam search with larger beam size (ablation) | Info-Gain Sampler · MATH-500 | cumulative uncertainty | No clear effect |
Closest papers by meaning
- Accelerated Sampling from Masked Diffusion Models via Entropy Bounded Unmasking (NeurIPS 2025 · similarity 0.92)
- Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusions (ICML 2025 · similarity 0.91)
- Learn from Your Mistakes: Self-Correcting Masked Diffusion Models (NeurIPS 2026 · similarity 0.90)
- Masked Diffusion Models are Secretly Time-Agnostic Masked Models and Exploit Inaccurate Categorical Sampling (ICLR 2025 · similarity 0.90)
- Beyond Masked and Unmasked: Discrete Diffusion Models via Partial Masking (NeurIPS 2025 · similarity 0.90)
- Attention-Discounted Adaptive Sampler for Masked Diffusion Language Models (NeurIPS 2026 · 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.
