Diffusion large language models (D-LLMs) have emerged as a promising alternative to auto-regressive models due to their iterative refinement capabilities. However, hallucinations remain a critical issue that hinders their reliability. To detect hallucination responses from model outputs, token-level uncertainty, such as entropy, has been widely used to indicate potential factual errors. Nevertheless, unlike auto-regressive models that generate tokens sequentially, D-LLMs generate fixed-length sequences simultaneously, where only a small subset of tokens is informative for hallucination detection. Thus, aggregating uncertainty over all tokens can be suboptimal. Moreover, the evolution trend of uncertainty throughout the diffusion process can also provide valuable signals, highlighting the necessity of modeling its denoising dynamics for hallucination detection. In this paper, we propose DynHD, which bridges these gaps from both spatial (token sequence) and temporal (denoising dynamics) perspectives. To handle the information density imbalance across tokens, we propose a semantic-aware evidence construction module that extracts hallucination-indicative signals by removing task-invariant structural tokens and emphasizing the uncertainty of the remaining informative tokens. To model denoising dynamics for hallucination detection, we introduce a reference evidence generator that learns the expected evolution trajectory of uncertainty evidence, along with a deviation-based hallucination detector that makes predictions by measuring the discrepancy between the observed and reference trajectories. Extensive experiments demonstrate that DynHD consistently outperforms state-of-the-art baselines while achieving higher efficiency across multiple benchmarks and backbone models. The code is available at: https://github.com/qyy11-com/DynHD.
Figures & tables
Figure 1: Visualization of spatial uncertainty distribution during decoding. Tokens exhibiting the highest entropy spikes serve as the primary indicators of factual instability. On the contrary, structural tokens provide limited cues for hallucination detection.
Figure 2: Entropy dynamics under spatial filtering. (a) Mean trajectories fail to distinguish between samples due to information density imbalance. (b) & (c) Filtered trajectories (top- K ) reveal distinct convergence: factual samples decay monotonically, while hallucinated ones exhibit late-stage stagnation or rebound. Differences in decay rates between HotpotQA (b) and TriviaQA (c) reflect varying task-specific denoising dynamics.
Figure 3: Framework overview of DynHD, which constructs a semantic-aware evidence trajectory from uncertainty, and detects hallucination via learning the deviation between the reference and observed trajectories.
Model
Method
TriviaQA
HotpotQA
CSQA
Avg
128
64
128
64
128
64
LLaDA-8B-Instruct
Output-based Methods
Perplexity
50.4
47.6
49.3
51.2
65.6
65.0
54.9
LN-Entropy
54.6
53.5
54.8
54.7
64.6
64.4
57.8
Semantic Entropy
68.9
67.3
57.6
53.8
44.1
43.9
55.9
Table 1: AUROC(%) comparison of hallucination detection methods on two D-LLMs across three QA datasets. The highest score is bolded and the second highest is underlined .
Figure 4: Pareto-frontier of performance vs. efficiency.
Figure 5: (a) Visualization of attention weights ωt on different steps. (b) Top- k entropy trajectories of representative samples on HotpotQA ( T=128 ). (c) Performance comparison of different D-LLM remasking strategies.
Ablation Variant
TriviaQA
HotpotQA
CSQA
128
64
128
64
128
64
LLaDA-8B-Instruct
DynHD (Full Framework)
86.6
86.2
84.3
85.2
81.5
84.2
Semantic-aware Evidence Construction Components
w/o Token Filtering
74.8
73.6
75.5
78.1
74.1
76.6
w/o Mean Uncertainty ( μavg )
78.6
81.1
79.0
80.2
78.5
77.1
Table 3: Ablation study in AUROC(%).
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Empirical analysis of the distinguishability of statistical evidence trajectories. (a) shows the enhanced separation after structure-aware filtering, while (b) and (c) demonstrate the consistent morphological divergence between factual and hallucinatory samples across datasets.
Variable
Search Range
LRstage1
{3×10−4,1×10−3}
LRstage2
{1×10−4,3×10−4,1×10−3}
WDstage1
{0,0.01,0.05}
WDstage2
{0,1×10−4,1×10−3}
λ1 (Path)
[0.0,0.4] (step 0.05 )
λ2 (Rebound)
[0.0,0.4] (step 0.05 )
Appendix
Table 4: Search space of key hyperparameters in DynHD.
Dataset
Type
Question & Response
HotpotQA
Factual
Q: Are Manhattan West and Singer Building both projects in New York?
R: Yes .
Hallucination
Q: What State has a Disney Resort & Spa that is a beachside hotel, resort and vacation destination offering complimentary children’s activities and programs and that Djuan Rivers was a General Manager at?
R: The Disney Resort and Spa where Djuan Rivers served as General Manager is located in Florida .
TriviaQA
Factual
Q: Where did the Shining Path terrorists operate?
R: The Shining Path terrorists operated in Peru .
Appendix
Table 5: Examples of factual and hallucinated samples. Each dataset displays a Factual sample followed by a Hallucination. Factual samples exhibit consistent convergence, whereas hallucinations trigger detectable instability in denoising dynamics.
Figure 7: AUROC heatmap for varying λ1 (Path) and λ2 (Rebound) on TriviaQA. The peak region validates the synergistic effect of both dynamical regularizers.
Figure 8: Empirical denoising dynamics of at (observed) and a^t (reference) at 64 steps on HotpotQA.
Diffusion large language models generate text through iterative denoising, exposing hidden trajectories that may contain reliability signals beyond the final output. We propose HIVE, which compresses trajectory hidden states, selects informative step-layer evidence, and conditions a verifier through continuous prefix embeddings to produce a hallucination score and structured diagnostics. Across two D-LLMs and three QA benchmarks, HIVE outperforms eight established baselines and a verifier-backbone-matched text-only control in all six settings. Relative to text-only verification, hidden-evidence conditioning improves AUROC by 1.73--4.60 points and AUPRC by 1.10--3.62 points, with average gains of 3.15 and 2.28 points, respectively. Ablations, evidence interventions, and cross-dataset transfer further support the complementary value of fine-grained hidden trajectory evidence.
Guoshenghui Zhao, Tan Yu, Weijie Zhao
Rochester Institute of Technology · NVIDIA Corporation
While Diffusion Large Language Models (dLLMs) have emerged as a promising non-autoregressive paradigm comparable to autoregressive (AR) models, their faithfulness, specifically regarding hallucination, remains largely underexplored. To bridge this gap, we present the first controlled comparative study to evaluate hallucination patterns in dLLMs. Our results demonstrate that current dLLMs exhibit a higher propensity for hallucination than AR counterparts controlled for architecture, scale, and pre-training weights. Furthermore, an analysis of inference-time compute reveals divergent dynamics: while quasi-autoregressive generation suffers from early saturation, non-sequential decoding unlocks potential for continuous refinement. Finally, we identify distinct failure modes unique to the diffusion process, including premature termination, incomplete denoising, and context intrusion. Our findings underscore that although dLLMs have narrowed the performance gap on general tasks, their distinct hallucination mechanisms pose a critical challenge to model reliability. Our code is available at https://github.com/ZeroLoss-Lab/Lost-in-Diffusion
Zhengnan Guo, Fei Tan
East China Normal University · Zhejiang University of Technology
Large Language Diffusion Models (LLDMs) are emerging as an alternative to autoregressive models, offering faster inference through higher parallelism. Similar to autoregressive LLMs, they remain prone to hallucinations, making reliable uncertainty quantification (UQ) crucial for safe deployment. However, existing UQ methods are fundamentally misaligned with this new paradigm: they assume autoregressive factorization or use expensive repeated sampling, negating the efficiency of LLDMs. In this work, we present the first systematic study of UQ for LLDMs and propose lightweight, zero-shot uncertainty signals derived from the iterative denoising process, leveraging intermediate generations, token remasking dynamics, and denoising complexity. We further adapt a state-of-the-art UQ method to LLDMs by combining masked diffusion likelihoods with trajectory-based semantic dissimilarity. We prove that expected trajectory dissimilarity lower bounds the masked diffusion training objective, which motivates its usage as an uncertainty score. Comprehensive experiments across three tasks, eight datasets, and two models show that our method achieves a great cost-performance trade-off: it approaches the strongest sampling-based baselines while incurring up to 100x lower computational overhead. Our work demonstrates that LLDMs can deliver both fast inference and reliable hallucination detection simultaneously.