Research map / Large language models / Language, safety and interpretability
Diffusion LLMs Can Do Faster-Than-AR Inference via Discrete Diffusion Forcing
Xu Wang, Chenkai Xu, Yijie Jin, Jiachun Jin, Hao Zhang, Kai Yu, Zhijie Deng
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 |
|---|---|---|---|
| apply D2F to LLaDA-Instruct | LLaDA-Instruct-8B · MBPP 3-shot | tokens per second: 0.9 -> 47.6 | Improved |
| apply D2F to LLaDA-Instruct | LLaDA-Instruct-8B · GSM8K 4-shot | score: 77.4 -> 77.3 | No clear effect |
| apply D2F to LLaDA-Instruct | LLaDA-Instruct-8B · MATH 4-shot | tokens per second: 21.1 -> 90.2 | Improved |
| apply D2F to Dream-Base | Dream-Base-7B · GSM8K-CoT 8-shot | tokens per second: 9.5 -> 91.2 | Improved |
| apply D2F to Dream-Base | Dream-Base-7B · GSM8K-CoT 8-shot | score: 75.0 -> 77.6 | Improved |
| add pipelined parallel decoding on top of KV cache (ablation) | D2F-LLaDA · GSM8K 4-shot | tokens per second: 17.5 -> 52.5 | Improved |
| add pipelined parallel decoding on top of KV cache (ablation) | D2F-Dream · MBPP 3-shot | tokens per second: 40.5 -> 105 | Improved |
| add pipelined parallel decoding on top of KV cache (ablation) | D2F-Dream · MATH 4-shot | score: 35.9 -> 35.4 | No clear effect |
| use D2F progressive block noising instead of random per-block noise (ablation) | D2F-Dream · MBPP 3-shot | score: 49.6 -> 54.6 | Improved |
| use D2F training instead of random noise per block (ablation) | D2F-Dream · GSM8K 4-shot (tau_act 0.95) | score: 77.1 -> 77.2 | No clear effect |
| use dual-state pipeline (tau_add 0.7 < tau_act) instead of tau_add = tau_act (ablation) | D2F-Dream · GSM8K 4-shot (tau_act 0.85) | score: 72.6 -> 74.2 | Improved |
| lower tau_add from 0.7 to 0.1 (ablation) | D2F-Dream · GSM8K 4-shot (tau_act 0.85) | score: 74.2 -> 75.0 | Improved |
Closest papers by meaning
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.93)
- SimSD: Simple Speculative Decoding in Diffusion Language Models (NeurIPS 2026 · similarity 0.92)
- DAWN: Dependency-Aware Fast Inference for Diffusion LLMs (NeurIPS 2026 · similarity 0.91)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.91)
- Beyond Autoregression: Fast LLMs via Self-Distillation Through Time (ICLR 2025 · similarity 0.91)
- Fast-dLLM++: Fr\'{e}chet Profile Decoding for Faster Diffusion LLM Inference (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.
