Diffusion language models can answer before reasoning is complete

A study of masked diffusion language models finds a failure mode in reasoning tasks: the model can commit to the final answer before enough of the reasoning region has been filled in. The authors log every token commitment during LLaDA-8B decoding on GSM8K.

In unconstrained decoding, the final answer appears 15 to 24 percent of the way through the trajectory while much of the reasoning remains masked. As the canvas grows, outputs collapse to answer-only responses on up to 90 percent of problems.

The paper argues the issue is not simply the model wanting to stop early. It is about reachability: whether the sampler is allowed to act on end-of-sequence beliefs at distant positions. For diffusion-style LLMs, decoding order may be as important as model size.