Draft Refinement Language Models 以草稿精煉實現並行推理
Breaking Diffusion: Parallel Reasoning with Draft Refinement
Draft Refinement Language Models(DRLMs)提出一種並行推理方法,先生成完整草稿,再反覆精煉,而非逐個生成 token 或採用擴散反向過程。
We introduce Draft Refinement Language Models (DRLMs), a new approach to parallel reasoning that generates a complete draft and iteratively refines it, rather than generating tokens sequentially or following a diffusion reverse process.
DRLM uses a single denoiser to propose parallel revisions and evaluate the resulting drafts. We derive a cavity-energy score that measures how well each token is supported by its surrounding context, allowing the model to retain better-scoring drafts without a separate verifier or scoring model.
Across Countdown, Sudoku-Extreme, and Maze-Hard, DRLM outperforms autoregressive and discrete diffusion baselines. With just 6M parameters, it achieves 44.4% on Countdown-5, 93.9% on Sudoku-Extreme, and 73.0% on Maze-Hard, outperforming the strongest diffusion baselines by 34.1, 28.4, and 33.7 percentage points, respectively.
Our results highlight complete-draft refinement as a promising alternative to conventional diffusion decoding, enabling reasoning through parallel revision and explicit evaluation of candidate solutions.
來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co