Parallel Refinement Instead of Diffusion: DRLMs Hit 93.9% on Sudoku-Extreme

A

New Parallel Reasoning Method: Generate and Refine Instead of Diffuse

Microsoft Research has introduced Draft Refinement Language Models (DRLMs), an innovative approach to modeling complex reasoning. The method departs radically from the two dominant paradigms: it follows neither the sequential token-by-token logic of autoregressive models nor the reverse-pass logic of diffusion. Instead, DRLMs generate a complete draft answer in one shot, then iteratively refine it by proposing and evaluating parallel corrections.

The core idea is simple: the final answer emerges from repeatedly refining the whole, rather than from a sequence of tiny token-level choices. At each iteration, the model considers the entire context — the original question and the current draft — simultaneously, which allows it to make more informed decisions about what to fix.

Architecture: A Single Denoiser and the Cavity-Energy Score

DRLM relies on a single denoiser — a module trained to reconstruct each token from uniformly corrupted drafts. This is dual-purpose training: the model learns both to preserve correct tokens and to fix erroneous ones. The denoiser's predictions directly determine the proposed corrections without running a reverse diffusion transform — this is what fundamentally sets DRLM apart from classic discrete diffusion models.

The central innovation is the cavity-energy score, a metric for evaluating draft quality computed in a single forward pass with no separate verifier model. Cavity-energy measures how well each token is supported by its surrounding context. The authors use a well-known data-corruption process to correct for self-token evidence — the influence of a token on its own score — yielding a clean measure of contextual support for each position.

The refinement process is strictly iterative: (1) based on the current draft, propose corrections for all positions in parallel; (2) compute the cavity-energy of the completed candidate in its refined context (a single forward pass); (3) keep the correction only if the new cavity-energy value is lower than before.

Results on Three Complex Reasoning Benchmarks

The authors tested DRLM on three tasks requiring deep combinatorial reasoning. All tests were run with a 6M-parameter model:

• Countdown-5: DRLM achieved 44.4% accuracy, 34.1 percentage points above the best discrete diffusion baseline (10.3%) • Sudoku-Extreme: DRLM solved 93.9% of puzzles, beating the baseline by 28.4 points (baseline: 65.5%) • Maze-Hard: DRLM scored 73.0%, 33.7 points above the baseline (baseline: 39.3%)

Gains of 28 to 34 percentage points represent a fundamental improvement. Importantly, DRLM requires no separate scoring model — cavity-energy is computed in parallel with the main model within a single forward pass.

Why It Matters

Full draft refinement is a viable and effective alternative to both reverse diffusion processes and autoregressive decoding. This is especially critical for complex reasoning tasks, where a model must coordinate choices across multiple variables simultaneously.

Parallel refinement lets the model explicitly evaluate entire candidates and select those best supported by context — a more honest measure of quality than choosing one token at a time, sequentially.

Authors: Jaesang Yu, Seoyeon Byun, Yun Wang, Yan Lu, Juho Kim

Organization: Microsoft Research

Publication date: October 19, 2026

Model size: 6M parameters