New LogicDiff method improves reasoning in masked diffusion language models by unmasking logical tokens in dependency order rather than by confidence.

Standard masked diffusion language models (MDLMs) defer high-entropy logical connectives during generation, causing reasoning performance to degrade significantly

LogicDiff uses a lightweight 4.2M parameter classification head (0.05% of base model size) to predict logical roles of masked tokens with 98.4% accuracy

The method implements logic-ordered unmasking: premises first, then connectives, then derived steps, then conclusions, replacing traditional confidence-based strategies

Inference-time approach requires no modification to the underlying model parameters, making it practical for deployment on existing masked diffusion models