Research map / Image generation / Fast diffusion sampling
Locally Coherent Parallel Decoding in Diffusion Language Models
Michael Hersche, Nicolas Menet, Ronan Tanios, Abbas Rahimi
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 CoDiLA local AR block to Dream-Coder (K=B=4) | Dream-Coder-Instruct-7B · HumanEval | Pass@1: 24.4 -> 51.8 | Improved |
| add CoDiLA local AR block to Dream-Coder (K=B=8) | Dream-Coder-Instruct-7B · HumanEval | Pass@1: 5.5 -> 29.9 | Improved |
| add CoDiLA to Dream-Coder | Dream-Coder-Instruct-7B · HumanEval syntax errors | syntax error rate: 70 -> 16 | Improved |
| use soft-conditioning instead of Top-1 (ablation) | CoDiLA (B=4) · HumanEval | Pass@1: 36.6 -> 51.8 | Improved |
| use soft-conditioning instead of Top-1 (ablation) | CoDiLA (B=4) · MBPP | Pass@1: 33.5 -> 55.6 | Improved |
| use soft-conditioning instead of Top-5 (ablation) | CoDiLA (B=4) · HumanEval | Pass@1: 56.1 -> 51.8 | Mixed |
| add boundary tokens around soft embeddings (ablation) | CoDiLA (B=4) · NELBO | NELBO: 15.5 -> 13.6 | Improved |
| increase block size B (ablation) | CoDiLA · training loss | loss | Improved |
| scale AR model to 1.7B or 4B (ablation) | CoDiLA · HumanEval/MBPP | Pass@1 | No clear effect |
| extend candidate scope from 10 to 50 blocks (ablation) | CoDiLA (B=4) · HumanEval | Pass@1: 51.8% -> 42.1% | Hurt |
| use global scope with EOS penalty instead of local scope (ablation) | CoDiLA · code benchmarks | Pass@1 | No clear effect |
| decode one block instead of two blocks in parallel (B=8) (ablation) | CoDiLA · HumanEval | Pass@1: 22.3 -> 39.3 | Improved |
Closest papers by meaning
- Mean-Field Parallel Decoding for Discrete Diffusion Language Models (NeurIPS 2026 · similarity 0.92)
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.92)
- Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models (NeurIPS 2026 · similarity 0.91)
- Dynamically Structured Diffusion Language Model Decoding via Bayesian Inference (NeurIPS 2026 · similarity 0.91)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.91)
- SimSD: Simple Speculative Decoding in Diffusion Language Models (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.
