Reasoning or Rationalization? Probing the Logic of Diffusion Models

ACL ARR 2026 January Submission7155 Authors

Keywords: fact verification, diffusion language models, reasoning, mdlm, dllm

Abstract: Unlike traditional autoregressive models which rely on causal reasoning to derive answers, Masked Diffusion Language Models (MDLMs) refine all sequence positions simultaneously, raising questions about the necessity of explicit reasoning steps in non-causal architectures. In this work, we investigate the dynamics of MDLM reasoning on fact verification. We observe that MDLMs typically converge on a verdict early in the generation process, treating it as a global anchor. Crucially, we find that enforcing a reasoning-first constraint via delayed verdict unmasking actively degrades performance due to refinement drift, where local noise overrides initially accurate judgments. Interventional experiments further reveal that MDLMs prioritize global sequence consistency over factual integrity, often hallucinating justifications to rationalize incorrect verdicts. Our findings suggest that for diffusion-based architectures, prolonged deliberation can be counter-productive, as it risks diluting accurate global priors with generated noise.

Paper Type: Short

Research Area: Interpretability and Analysis of Models for NLP

Research Area Keywords: fact checking, explanation faithfulness, counterfactual/contrastive explanations, chain-of-thought, logical reasoning, generative models

Contribution Types: Model analysis & interpretability

Languages Studied: English

Submission Number: 7155