Research map / Image generation / Fast diffusion sampling
Unveiling the Potential of Diffusion Large Language Model in Controllable Generation
Zhen Xiong, Yujun Cai, Zhecheng Li, Yiwei Wang
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 schema scaffolding (S3) instead of direct prompting (ablation) | LLaDA-1.5 · WikiBio, 8 steps | Schema Compliance: 0.312 -> 0.994 | Improved |
| add schema scaffolding (S3) (ablation) | LLaDA-1.5 · WikiBio, 32 steps | Schema Compliance: 0.792 -> 0.997 | Improved |
| add 3 few-shot examples to baseline (ablation) | LLaDA-1.5 · WikiBio, 8 steps | Schema Compliance: 0.312 -> 0.443 | Improved |
| add complete schema to the prompt (ablation) | LLaDA-1.5 · WikiBio, 16 steps | Schema Compliance: 0.576 -> 0.738 | Improved |
| add S3 instead of baseline (ablation) | LLaDA-1.5 · WikiBio, 8 steps | F1: 0.068 -> 0.130 | Improved |
| add few-shot examples (ablation) | LLaDA-1.5 · WikiBio, 8 steps | F1: 0.068 -> 0.068 | No clear effect |
| use vanilla scaffolding S2 (fixed mask slots) instead of baseline (ablation) | LLaDA-1.5 · WikiBio, 8 steps | Hallucination Rate: 0.404 -> 0.465 | Hurt |
| replace S2 by S3 with null placeholder for surplus slots (ablation) | LLaDA-1.5 · WikiBio, 8 steps | Hallucination Rate: 0.465 -> 0.340 | Improved |
| use vanilla S2 scaffolding (ablation) | LLaDA-1.5 · WikiBio | precision | Hurt |
| increase denoising steps from 8 to 32 (ablation) | baseline LLaDA-1.5 · WikiBio | Schema Compliance: 0.312 -> 0.792 | Improved |
| increase denoising steps (ablation) | LLaDA-1.5 content fidelity · WikiBio | content fidelity | No clear effect |
| add few-shot examples (ablation) | LLaDA-1.5 · WikiBio, 8 steps | Hallucination Rate: 0.404 -> 0.366 | Improved |
Closest papers by meaning
- Language Control Diffusion: Efficiently Scaling through Space, Time, and Tasks (ICLR 2024 · similarity 0.88)
- TRACE Back from the Future: A Probabilistic Reasoning Approach to Controllable Language Generation (ICML 2025 · similarity 0.87)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.87)
- The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models (ICML 2026 · similarity 0.87)
- Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo (ICLR 2025 · similarity 0.87)
- Efficient Test-Time Scaling via Hierarchical Search and Self-Verification for Discrete Diffusion Language Models (ICML 2026 · 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.
