Research map / Image generation / Fast diffusion sampling
Continuous Diffusion Models Can Obey Formal Syntax
Jinwoo Kim, Taylor Berg-Kirkpatrick, Loris D'Antoni
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 |
|---|---|---|---|
| use Diffinity instead of GPT2-Large constrained decoding | PLAID · Natural language benchmarks | constraint satisfaction (%): 20.5 -> 92.9 | Improved |
| use Diffinity instead of GPT2-Large Guidance | PLAID · JSON schema | constraint satisfaction (%): 77.0 -> 68.4 | Hurt |
| use Diffinity instead of LLaDA+DINGO | PLAID · JSON schema | pass@10 (%): 100.0 -> 91.4 | Hurt |
| add DINGO constrained decoding (ablation) | LLaDA · Natural language benchmarks | perplexity: 23.9 -> 222.8 | Hurt |
| add Diffinity guidance (scale 2.5) (ablation) | PLAID · Natural language benchmarks | perplexity: 61.6 -> 60.0 | No clear effect |
| add Diffinity guidance (scale 2.5) (ablation) | PLAID · Natural language benchmarks | fluency: 46.0 -> 35.8 | Hurt |
| use <json><EOS>.* padding instead of .*<json>.* padding (ablation) | Diffinity · JSON schema | constraint satisfaction (%): 68.4 -> 56.8 | Hurt |
| use <json><EOS>+ padding instead of .*<json>.* padding (ablation) | Diffinity · JSON schema | constraint satisfaction (%): 68.4 -> 33.6 | Hurt |
| replace regex-based expected-probability guidance with classifier guidance (full) (ablation) | Diffinity · Natural language benchmarks | constraint satisfaction (%): 92.9 -> 2.27 | Hurt |
| replace regex-based guidance with classifier guidance (ablation) | Diffinity · JSON schema | constraint satisfaction (%): 68.4 -> 0.00 | Hurt |
| use Diffinity instead of GPT2-Small-GCD for sentence prefix distribution | PLAID · The versus It prefix regex | The:It sample ratio: 1.2 -> 0.03 | Improved |
| use Diffinity instead of GPT2-Large for perplexity on Prefix | PLAID · Natural language: Prefix | perplexity: 401.7 -> 66.5 | Improved |
Closest papers by meaning
- Diffusion Beats Autoregressive in Data-Constrained Settings (NeurIPS 2025 · similarity 0.91)
- Constrained Decoding for Diffusion Language Models via Efficient Inference over Finite Automata (NeurIPS 2026 · similarity 0.91)
- Constrained Discrete Diffusion (NeurIPS 2025 · similarity 0.91)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.91)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.91)
- Flow Map Language Models: One-step Language Modeling via Continuous Denoising (NeurIPS 2026 · similarity 0.91)
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.
