The Confidence Shortcut: A Reasoning Failure Mode of Masked Diffusion Models
Identifies a fundamental misalignment in confidence-based decoding for autoregressive-free generation, relevant to improving reasoning in diffusion-based language models.
Masked diffusion language models using confidence-based decoding produce high-confidence errors on complex reasoning tasks like multi-digit addition, with confidence-aligned training amplifying error rates by an order of magnitude, while random masking preserves reasoning accuracy.
Masked diffusion language models (MDMs) uniquely support any-order generation, with confidence-based decoding currently serving as the de facto standard inference policy. To optimize for this, recent training schemes attempt to align training mask patterns directly with those observed during generation. However, we argue that confidence-based decoding is inherently misaligned with the logical-flow trajectories required for complex reasoning, and that confidence-aligned training actively entrenches this misalignment. We make this concrete using multi-digit addition, where the decoding strategy prematurely predicts locally easy digits before resolving their long-range dependencies, producing high-confidence errors on challenging inputs. While traditional random masking keeps the failure rate low on this challenging tail, confidence-aligned training amplifies the error rate by an order of magnitude. Across five distinct reasoning tasks, this same pattern emerges with task-dependent severity: confidence-based decoding induces failures on highly complex inputs, and confidence-aligned training exacerbates them. In contrast, random masking -- despite its perceived inefficiency -- robustly preserves the reasoning-trajectory conditionals essential for solving the challenging tail.