Efficient Self-Evaluation for Diffusion Language Models via Sequence Regeneration

TL;DR

This paper introduces DiSE, a confidence estimation method for diffusion large language models that assesses output quality through sequence regeneration probabilities, enabling better self-evaluation and adaptive generation.

Contribution

The paper presents DiSE, a novel self-evaluation technique for dLLMs, and a flexible-length generation framework that improves quality assessment and output control.

Findings

DiSE correlates with semantic coherence and answer accuracy.

DiSE enables reliable uncertainty quantification.

The flexible-length generation adapts sequence length based on self-assessment.

Abstract

Diffusion large language models (dLLMs) have recently attracted significant attention for their ability to enhance diversity, controllability, and parallelism. However, their non-sequential, bidirectionally masked generation makes quality assessment difficult, underscoring the need for effective self-evaluation. In this work, we propose DiSE, a simple yet effective self-evaluation confidence quantification method for dLLMs. DiSE quantifies confidence by computing the probability of regenerating the tokens in the entire generated sequence, given the full context. This method enables more efficient and reliable quality assessment by leveraging token regeneration probabilities, facilitating both likelihood estimation and robust uncertainty quantification. Building upon DiSE, we further introduce a flexible-length generation framework, which adaptively controls the sequence length based on the model's self-assessment of its own output. We analyze and validate the feasibility of DiSE from the perspective of dLLM generalization, and empirically demonstrate that DiSE is positively correlated with both semantic coherence and answer accuracy. Extensive experiments on likelihood evaluation, uncertainty quantification, and flexible-length generation further confirm the effectiveness of the proposed DiSE.