Efficient Self-Evaluation for Diffusion Language Models via Sequence Regeneration

Researchers propose DiSE, a self-evaluation method for diffusion large language models (dLLMs) that quantifies confidence by computing token regeneration probabilities. The method enables more efficient quality assessment and introduces a flexible-length generation framework that adaptively controls sequence length based on the model's self-assessment.

• →DiSE provides a simple yet effective confidence quantification method for diffusion large language models through sequence regeneration.
• →The method leverages token regeneration probabilities to enable likelihood estimation and uncertainty quantification.
• →A flexible-length generation framework adaptively controls sequence length based on model self-assessment.
• →DiSE shows positive correlation with both semantic coherence and answer accuracy in experimental validation.
• →The approach addresses quality assessment challenges in non-sequential, bidirectionally masked generation of dLLMs.