Research map / Image generation / Fast diffusion sampling
Swordsman: Entropy-Driven Adaptive Block Partition for Efficient Diffusion Language Models
Yu Zhang, Xinchen Li, Jialei Zhou, Hongnan Ma, Zhongwei Wan, Yiwei Shi, Duoqian Miao, Qi Zhang, Longbing Cao
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 fixed-size blocks with entropy-driven adaptive block partition (ablation) | Swordsman on LLaDA-8B-Instruct · GSM8K (no cache) | accuracy: 77.56 -> 81.43 | Improved |
| add block-wise decoding to full-diffusion LLaDA (ablation) | LLaDA-8B-Instruct · GSM8K | accuracy: 29.04 -> 77.40 | Improved |
| add confidence-based parallel decoding to block-wise decoding (ablation) | LLaDA-8B-Instruct · GSM8K | throughput (TPS): 8.39 -> 42.36 | Improved |
| use dynamic threshold (tau_init 0.9) instead of static threshold 0.9 (ablation) | Swordsman on LLaDA-8B-Instruct · GSM8K (no cache) | latency (s): accuracy 81.45 vs 81.43; latency reduced by 2.49s | Improved |
| lower initial dynamic threshold to 0.5 (ablation) | Swordsman on LLaDA-8B-Instruct · GSM8K (no cache) | accuracy: 81.43 -> 68.46 | Hurt |
| raise initial dynamic threshold to 1.0 (ablation) | Swordsman on LLaDA-8B-Instruct · GSM8K (no cache) | latency (s): about 2.4x higher latency | Hurt |
| use Swordsman instead of Fast-dLLM under dual cache | LLaDA-1.5 · HumanEval (dual cache) | accuracy: 35.59 -> 43.90 | Improved |
| use Swordsman instead of Fast-dLLM under dual cache | LLaDA-8B-Instruct · GSM8K (dual cache) | accuracy: 75.21 -> 81.50 | Improved |
| use Swordsman instead of Fast-dLLM under dual cache | LLaDA-8B-Instruct · GSM8K speed | throughput (TPS): 72.12 -> 73.74 | Improved |
| use Swordsman instead of Fast-dLLM on MATH under dual cache | LLaDA-1.5 · MATH (dual cache) | accuracy: 33.26 -> 35.30 | Improved |
| use Swordsman instead of Fast-dLLM on MBPP under dual cache | LLaDA-8B-Instruct · MBPP (dual cache) | accuracy: 13.60 -> 13.60 | No clear effect |
| use Swordsman instead of Fast-dLLM on MATH under dual cache | Dream-v0-Base-7B · MATH (dual cache) | accuracy: 38.34 -> 38.58 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Fast-dLLM — Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding (ICLR 2026) · better on 11 of 12 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.91)
- PACE-dLLM: Elastic Block Decoding via Confidence Cliff Estimation for Diffusion Language Models (NeurIPS 2026 · similarity 0.91)
- Confidence-Based Decoding is Provably Efficient for Diffusion Language Models (NeurIPS 2026 · similarity 0.91)
- Encoder-Decoder Diffusion Language Models for Efficient Training and Inference (NeurIPS 2025 · similarity 0.90)
- Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models (ICLR 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.
