Breaking Diffusion: Parallel Reasoning with Draft Refinement
We introduce Draft Refinement Language Models (DRLMs), which use a single denoiser to propose parallel revisions and evaluate a candidate draft.
ProofPaper ↗
Key points
- Drafting an answer allows a reasoning model to reconsider each decision in the context of the others.
- Yet repeated diffusion denoising does not ensure that individually plausible token updates form a better answer together.
- Across Countdown, Sudoku-Extreme, and Maze-Hard, our method outperforms autoregressive and discrete diffusion baselines.
- The results establish complete-draft parallel refinement as an effective alternative to reverse diffusion for complex reasoning.
Sources (2)
- [1]Breaking Diffusion: Parallel Reasoning with Draft RefinementarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 03:02 PM
We introduce Draft Refinement Language Models (DRLMs), which use a single denoiser to propose parallel revisions and evaluate a candidate draft.
Drafting an answer allows a reasoning model to reconsider each decision in the context of the others.
- [2]Breaking Diffusion: Parallel Reasoning with Draft RefinementHugging Face Daily Papers · Oct 19, 12:00 AM · same content
Extractive summary: sentences quoted from the sources.
Before this
- Oct 16, 2026ReLaDrift: Recurrent Language Generation via Drifting for Test-Time Scaling
- Oct 16, 2026Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline
- Oct 16, 2026A Conditional No-Hair Theorem for Transformers: How Language Models Forget Their Prompt
- Oct 16, 2026Towards Looped Models Done Right, Part I: Topology, Input Injection, Recurrent-State Design
- Oct 16, 2026HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents
- Oct 16, 2026Language Models Learn to Improve Language Models