Research map / Architectures and efficiency / Attention and transformers
Fast and Accurate Causal Parallel Decoding using Jacobi Forcing
Lanxiang Hu, Siqi Kou, Yichao Fu, Samyam Rajbhandari, Tajana Rosing, Yuxiong He, Zhijie Deng, Hao Zhang
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 linear progressive noise schedule instead of random (window 8) (ablation) | JF-LLM · HumanEval (10k instances) | pass@1 accuracy: 82.9 -> 84.7 | Improved |
| use reverse progressive noise schedule instead of linear progressive (ablation) | JF-LLM · HumanEval (10k instances) | pass@1 accuracy: 84.7 -> < 80.0 (window 8) | Hurt |
| increase progressive schedule window size from 8 to 16 (ablation) | JF-LLM · HumanEval (10k instances) | pass@1 accuracy: 84.7 -> 81.7 | Hurt |
| use full noise-conditioned mask rather than intra-window clean context (ablation) | JF-LLM · HumanEval (100k examples) | speedup: 1.9× -> 3.6× | Improved |
| add multi-block decoding with rejection recycling | JF-LLM · HumanEval (A100) | speedup: 3.86× -> 3.97× | Improved |
| add multi-block decoding with rejection recycling | JF-LLM · MATH (A100) | speedup: 3.65× -> 3.68× | No clear effect |
| add second round of progressive distillation with larger block size | JF-LLM · HumanEval | speedup: further 20% speedup, minor performance degradation | Improved |
| increase number of decoding blocks beyond 2 (ablation) | JF-LLM (MR) · HumanEval | speedup | No clear effect |
| replace AR-adapted dLLMs with Jacobi Forcing causal parallel decoder | Qwen2.5-Coder-7B-Instruct · HumanEval | speedup: 1.77× (D2F) -> 3.86× | Improved |
| Jacobi Forcing training vs lossless AR model | Qwen2.5-Coder-7B-Instruct · HumanEval | pass@1 accuracy: 87.8 -> 83.5 | Hurt |
| Jacobi Forcing training vs EAGLE-3 speculative decoding | Qwen2.5-Coder-7B-Instruct · GSM8K | speedup: 3.16× (EAGLE-3) vs 3.50× (JF-LLM) | Mixed |
Closest papers by meaning
- SimSD: Simple Speculative Decoding in Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models (NeurIPS 2026 · similarity 0.90)
- CLLMs: Consistency Large Language Models (ICML 2024 · similarity 0.90)
- Mean-Field Parallel Decoding for Discrete Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- Fast-dLLM++: Fr\'{e}chet Profile Decoding for Faster Diffusion LLM Inference (NeurIPS 2026 · similarity 0.90)
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.89)
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.
