Research map / Image generation / Fast diffusion sampling
Loopholing Discrete Diffusion: Deterministic Bypass of the Sampling Wall
Mingyu Jo, Jaesik Yoon, Justin Deschenaux, Caglar Gulcehre, Sungjin Ahn
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 loopholing (deterministic latent pathway) | MDLM · LM1B | test perplexity: 27.60 -> 25.95 | Improved |
| add loopholing | MDLM · OpenWebText | test perplexity: 23.05 -> 21.90 | Improved |
| add loopholing | UDLM · OpenWebText generation (1024 steps) | generative perplexity: 73.95 -> 28.76 | Improved |
| add loopholing | MDLM · OpenWebText generation (1024 steps) | generative perplexity: 108.94 -> 49.13 | Improved |
| add loopholing | MGDM 85M · Game of 24 | success rate (%): 47 -> 63 | Improved |
| add loopholing | MGDM 85M · Countdown 4 | success rate (%): 86.5 -> 94.4 | Improved |
| add loopholing | UDLM · zero-shot LM1B (OWT-trained) | perplexity: 81.41 -> 79.60 | Hurt |
| add loopholing | MDLM · zero-shot LM1B (OWT-trained) | perplexity: 66.73 -> 69.53 | Hurt |
| add loopholing | MDLM · LAMBADA | accuracy (%): 40.46 -> 52.40 | Improved |
| add loopholing | MDLM · multiple-choice downstream tasks | accuracy: 36.49 -> 36.03 | No clear effect |
| add loopholing to autoregressive model | AR · unconditional generation | generative perplexity: 34.40 -> 34.21 | No clear effect |
| add MLP to deterministic path (ablation) | LDDM-M · LM1B | test perplexity: 25.95 -> 25.87 | No clear effect |
Closest papers by meaning
- Looped Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution (ICML 2024 · similarity 0.88)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.87)
- The Diffusion Duality (ICML 2025 · similarity 0.87)
- Constrained Discrete Diffusion (NeurIPS 2025 · similarity 0.87)
- Think while You Generate: Discrete Diffusion with Planned Denoising (ICLR 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.
