When Confidence Misleads: Suffix Anchoring and Anchor-Proximity Confidence Modulation for Diffusion Language Models

Researchers found a failure mode in fully non-autoregressive diffusion language model decoding: high-confidence EOT tokens and overconfidence near a suffix anchor can cause premature or incomplete outputs.

They proposed Suffix-Anchored Confidence Modulation, a training-free decoding method that adds a short suffix anchor and rebalances confidence near it.

The approach improved results on text reasoning, vision-language reasoning, and code generation benchmarks.

It preserves the parallel decoding advantages of diffusion models while reducing early token selection and output truncation.