DynHD: Detecting Hallucinations in Diffusion LLMs by Tracking Denoising Dynamics

Diffusion Large Language Models (D-LLMs) generate answers by iteratively denoising from a heavily masked sequence, unlike autoregressive models that produce tokens sequentially. This paradigm raises a new question: when a D-LLM hallucinates, should we only inspect the final output, or examine how the answer emerged step by step?

A Key Observation: Not All Tokens Are Equally Informative

D-LLMs typically generate on fixed-length sequences, so the output contains structural tokens such as padding and boundary markers. These tokens contribute little to factuality judgments but dilute aggregate uncertainty statistics. The authors found that averaging entropy across all tokens makes correct and hallucinated trajectories nearly indistinguishable (Figure 2).

After filtering out structural tokens and focusing on high-uncertainty tokens, a clear divergence appears: correct answers show steadily decreasing uncertainty, while hallucinated answers exhibit stagnation or rebound in later denoising steps (Figure 3). This indicates that hallucination signals reside not only in which tokens are uncertain, but in how those uncertainties evolve during denoising.

DynHD: Detecting Hallucinations via Dynamical Deviation Learning

Based on this observation, DynHD comprises two components:

Semantic-aware Evidence Construction – filters structural tokens from the fixed-length generation and extracts representative uncertainty information from the remaining tokens, producing a hallucination evidence vector for each denoising step.

Dynamical Deviation Learning – learns what a normal denoising trajectory should look like across different questions (a reference trajectory), then compares the actual trajectory against this reference. When the model's uncertainty dynamics deviate markedly — e.g., late-stage stagnation or rebound — DynHD flags the output as hallucinated.

The core idea is not to judge whether entropy at a single step is high, but whether the entire denoising process follows a normal, stable convergence pattern.

Experimental Results

Experiments were conducted on LLaDA-8B-Instruct and Dream-7B-Instruct using three QA datasets: TriviaQA, HotpotQA, and CommonsenseQA.

On LLaDA-8B-Instruct, DynHD achieves an average AUROC of 84.2% , substantially outperforming the trajectory-based baseline TraceDet ( 72.0% ).

On Dream-7B-Instruct, DynHD reaches an average AUROC of 84.3% .

In cross-dataset zero-shot evaluation, DynHD scores 72.9% average AUROC vs. TraceDet's 66.3% .

Regarding efficiency, DynHD does not require repeated sampling like some uncertainty methods; it directly uses the denoising trajectory already produced during a single D-LLM generation, striking a favorable performance-efficiency trade-off (Figure 5).

Conclusion

DynHD starts from a simple premise: for diffusion language models, the final answer is merely the last state of the generation process, while much hallucination-relevant information already appears in the preceding denoising trajectory. Experiments confirm that token-level information density differences and late-stage dynamical changes serve as strong signals for identifying D-LLM hallucinations. Modeling how an answer gradually forms — rather than only observing the final output — may offer a more natural perspective for reliability research on diffusion language models.

Paper: https://arxiv.org/abs/2603.16459

Code: https://github.com/qyy11-com/DynHD