All publications

COLM 2026 · Accepted

Locally Confident, Globally Stuck: The Quality-Exploration Dilemma in Diffusion Language Models

Liancheng Fang, Aiwei Liu, Henry Peng Zou, Yankai Chen, Enze Ma, Leyi Pan, Chunyu Miao, Wei-Chieh Huang, Xue Liu, Philip S. Yu

Overview

Low-confidence remasking can improve individual answers from diffusion language models while narrowing the range of reasoning paths they explore. This work explains that trade-off and develops an Independent Metropolis–Hastings sampler to balance generation quality with diversity. Experiments evaluate mathematical reasoning and code generation across MATH500, AIME24/25, HumanEval, and MBPP.

Resources