Research map / Image generation / Fast diffusion sampling
Stopping Computation for Converged Tokens in Masked Diffusion-LM Decoding
Daisuke Oba, Danushka Bollegala, Masahiro Kaneko, Naoaki Okazaki
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 |
|---|---|---|---|
| lock converged token positions (SureLock) in decoding | LLaDA-8B-Instruct · MT-Bench subset, Ngen=256 | algorithmic FLOPs ratio: 1.0 -> 0.496 | Improved |
| lock converged positions (SureLock) | LLaDA-8B-Instruct · MT-Bench single-turn | GPT-4o judge score: 4.8 -> 4.9 (Ngen=128) | No clear effect |
| lock converged positions (SureLock) | LLaDA-8B-Base · WikiText-103, Ngen=64, steps=64 | Gen.-PPL: 3.4813 -> 4.5722 | Hurt |
| lock converged positions (SureLock) | LLaDA-8B-Base · WikiText-103, Ngen=512, steps=512 | Gen.-PPL: 1.4658 -> 1.5248 | No clear effect |
| tighten KL threshold from 5e-3 to 5e-8 (ablation) | SureLock · WikiText-103, Ngen=64 | Gen.-PPL: 4.7596 -> 4.0033 | Improved |
| tighten KL threshold from 5e-2 to 5e-4 (ablation) | SureLock · MT-Bench subset, Ngen=64 | FLOPs ratio: 0.482 -> 0.547 | Hurt |
| combine SureLock with selection-based sparsification (ablation) | SureLock · MT-Bench, Ngen=256 | FLOPs ratio: 0.54 -> 0.28 | Improved |
| apply SureLock on HumanEval code generation | LLaDA-8B-Instruct · HumanEval (first 70) | Pass@1: 18.6 -> 18.6-20.3 | No clear effect |
| apply SureLock at small workload | LLaDA-8B-Instruct · runtime, Ngen=64, B=1 | tokens per second | No clear effect |
Closest papers by meaning
- MAGE: All-[MASK] Block Already Knows Where to Look in Block Diffusion LLM (NeurIPS 2026 · similarity 0.89)
- Accelerated Sampling from Masked Diffusion Models via Entropy Bounded Unmasking (NeurIPS 2025 · similarity 0.89)
- Mean-Field Parallel Decoding for Discrete Diffusion Language Models (NeurIPS 2026 · similarity 0.88)
- Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusions (ICML 2025 · similarity 0.88)
- Attention-Discounted Adaptive Sampler for Masked Diffusion Language Models (NeurIPS 2026 · similarity 0.87)
- Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs (NeurIPS 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.
