HIVE: Hidden-Evidence Verification for Hallucination Detection in Diffusion Large Language Models

2026-04-28

AI Summary

Existing approaches struggle to effectively detect hallucinations generated by diffusion-based large language models during their multi-step denoising process, as they rely solely on final outputs or coarse-grained trajectory statistics and overlook informative dynamic signals in intermediate hidden states. This work proposes HIVE, a novel framework that, for the first time, extracts and selects information-rich step-layer hidden evidence from the denoising trajectory. By leveraging a dual-stream representation and a learnable selection mechanism, HIVE injects this evidence into a verification language model as prefix embeddings, enabling fine-grained and interpretable hallucination detection. Evaluated on two diffusion models across three question-answering benchmarks, HIVE substantially outperforms eight strong baselines, achieving an AUROC of 0.9236 and an AUPRC of 0.9537.

Abstract

Diffusion large language models generate text through multi-step denoising, where hallucination signals may emerge throughout the trajectory rather than only in the final output. Existing detectors mainly rely on output uncertainty or coarse trace statistics, which often fail to capture the richer hidden dynamics of D-LLMs. We propose HIVE, a hidden-evidence verification framework that extracts compressed hidden evidence from denoising trajectories, selects informative step-layer evidence, and conditions a verifier language model on the selected evidence through prefix embeddings. HIVE produces both a continuous hallucination score from verifier decision logits and structured verification outputs, including hallucination types, evidence pairs, and short rationales. Across two D-LLMs and three QA benchmarks, HIVE consistently outperforms eight strong baselines and achieves up to 0.9236 AUROC and 0.9537 AUPRC. Ablation studies further confirm the importance of hidden-evidence conditioning, learned evidence selection, two-stream evidence representation, and step-layer embeddings. These results suggest that selected hidden evidence from denoising trajectories provides a stronger and more usable hallucination signal than output-only uncertainty or coarse trace statistics.

Problem

• hallucination detection
• diffusion large language models
• hidden evidence
• denoising trajectories
• verification

Innovation

• hidden-evidence verification
• diffusion LLMs
• denoising trajectory
• hallucination detection
• evidence selection