AI Paper Atlas

Research map / Large language models / Reasoning and chain of thought

DiffuReason: Enhancing Reasoning Ability for Diffusion Language Models via Monte Carlo Tree Search

YIPING SONG, Jinyu You, Zhiliang Tian, Jinsong Su, Minlie Huang, Chenping Hou

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DiffuReason, combining Monte Carlo Tree Search with diffusion language models to improve their reasoning accuracy while keeping the parallel-generation speed advantage over autoregressive MCTS.

Problem

Diffusion LLMs offer fast parallel generation but lag in accuracy on complex reasoning, while AR models with MCTS reason well but are slow due to serial token generation.

Approach

Applies Monte Carlo Tree Search to diffusion language models to add System 2-style search on top of parallel dLLM generation, aiming to get reasoning gains with lower latency than AR+MCTS. Details beyond the abstract are not available.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add lookahead simulation (ablation)DiffuReason (WeDLM) · MATH500accuracy (%): 64.6 -> 68.2Improved
add lookahead simulation (ablation)DiffuReason (WeDLM) · Gaokao2023accuracy (%): 46.4 -> 51.2Improved
add revision (ablation)DiffuReason (WeDLM) · MATH500accuracy (%): 63.0 -> 68.2Improved
add revision (ablation)DiffuReason (WeDLM) · Gaokao2023accuracy (%): 48.1 -> 51.2Improved
use block-wise sequential search instead of random node search (ablation)DiffuReason (LLaDA) · MATH500 subset (100 problems)overall accuracy (%): 9.4 -> 37.0Improved
use fast generation (32 steps) instead of fine generation (128 steps) (ablation)DiffuReason (LLaDA) · MATH500 subset (100 problems)overall accuracy (%): 39.0 -> 37.0, time 136.7 s -> 86.9 sMixed
use fast generation (32 steps) instead of 128 steps (ablation)DiffuReason (LLaDA) · MATH500 Level 5accuracy (%): 10.0 -> 5.0Hurt
use block length 64 instead of 32 (ablation)DiffuReason (LLaDA) · MATH500 subsetaccuracy (%): 28.0 -> 37.0Improved
use block length 64 instead of 128 (ablation)DiffuReason (LLaDA) · MATH500 subsetaccuracy (%): 30.0 -> 37.0Improved
apply DiffuReason search to LLaDA-8BLLaDA-8B · AIME2024accuracy (%): 6.7 -> 10.0Improved
apply DiffuReason search to WeDLM-8BWeDLM-8B · AIME2024accuracy (%): 3.3 -> 20.0Improved
apply DiffuReason search to WeDLM-8BWeDLM-8B · MATH500accuracy (%): 58.8 -> 68.2Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Experimental setup

Baselines: auto-regressive MCTS

Builds on: Monte Carlo Tree Search

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy