The Confidence Shortcut: A Reasoning Failure Mode of Masked Diffusion Models

In three linesMasked diffusion models (MDMs) with confidence-based decoding fail on complex reasoning tasks. Confidence-aligned training amplifies errors by an order of magnitude on multi-digit addition. Random masking better preserves the logical trajectories required for reasoning.

Summary generated by Claude — human-verified