acceptodds
Under review as a conference paper at ICLR 2027

Inducing Hallucinations in Video Large Language Models by Exploiting Spatiotemporal Corruption Sensitivity

Abstract

Video Large Language Models (Video-LLMs) have demonstrated strong capabilities in open-ended video understanding, yet remain susceptible to video manipulations. Existing video manipulations mainly aim to mislead Video-LLMs into producing wrong labels or predefined responses, leaving hallucinations, i.e., plausible responses containing multiple visually unsupported errors, unexplored. In this paper, we address this research gap by proposing VidHaze, the first approach to inducing hallucinations in Video-LLMs via video perturbations. Our key observation is a clear sensitivity gap between correct and hallucinated responses to diverse spatiotemporal video corruptions, where correct responses exhibit larger clean-to-corrupted distribution shifts. Building on this observation, VidHaze selects spatiotemporally corrupted videos as hard negatives and contrasts their next-token distributions with those of clean videos under a shared prefix, converting the resulting directional shifts into hallucination-oriented soft targets. Experiments on VATEX and MSRVTT datasets across four advanced Video-LLMs show that VidHaze achieves average sentence-level hallucination rates of 60.67%/34.27% in white-/black-box settings, while maintaining clean-response linguistic quality. Further analyses on when and why VidHaze is effective show its dependence on model family and dataset characteristics, and that its gains stem from increased hallucination density rather than response length.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Adversarial Perturbation, Video Large Language Models, Hallucination

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ja89OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents VidHaze, an adversarial perturbation approach to inducing hallucinations in Video-LLMs. Motivated by the observation that correct sentences are more sensitive to spatiotemporal corruptions than hallucinated ones, VidHaze selects a corrupted hard negative via Thompson Sampling and contrasts it with the clean video to build a hallucination-oriented soft target for PGD optimization. Experiments on VATEX and MSRVTT across four Video-LLMs show higher hallucination rates than adapted baselines in most settings.

Critical strengths
  1. The evaluation is extensive, covering four Video-LLMs, two datasets, and both white-box and black-box settings.
  2. Finding 3 is an interesting analysis. It shows with normalized metrics (HSR, HUR) that VidHaze increases hallucination density rather than merely expanding the generated response.
Critical weaknesses
  1. The motivation for deliberately increasing hallucinations is unclear, since even the latest Video-LLMs already hallucinate on clean videos (Table 1). The threat model, i.e., who would launch this attack and why, is not discussed.
  2. The VATEX results of Gemma4 and LLaVA-NV in Table 8 are identical to Table 1, including judge-dependent metrics (Qual., HSR, HUR), although the judge is changed. Please clarify.
  3. Eq. (17) is essentially VCD [1] in other words. Since s⁺ − s⁻ = ℓʰ − ℓᶜ, it reduces to softmax(2ℓʰ − ℓᶜ), i.e., VCD with the roles swapped. This component is also not ablated in Table 3.
  4. TVA already induces fluent hallucinations in Table 1, which contradicts Figure 1 and shows that hallucination induction in Video-LLMs is not a new problem.
  5. Simple baselines are missing, e.g., feeding the corrupted video directly, random noise under the same ε, or an untargeted attack.
  6. No human evaluation verifies the LLM judges.
  7. The gains in Table 2 are marginal (about 2.3 HSR points over Mirage on average), and no variance is reported.
Questions for authors

Why build the target from spatiotemporal hard negatives instead of the model’s second-best tokens on the original video?

1 comment
AuthorsOpenReview ↗

Looking Forward to Your Engagement in the Rebuttal Discussion

Thank you for your initial feedback. We believe we have thoroughly addressed your concerns through additional experiments and clarifications. We would appreciate your continued engagement in the discussion.

Reviewer_TVviOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies adversarial video perturbations that induce fluent but visually unsupported descriptions without prescribing a target response. Its motivating observation is that correct sentences exhibit larger changes in teacher-forced next-token distributions under selected spatiotemporal corruptions than hallucinated sentences. VidHaze selects an informative corrupted video, constructs a contrastive soft target from its distribution and the clean-video distribution, and optimizes a bounded pixel perturbation using position-dependent weights. Experiments on 100 videos each from VATEX and MSRVTT demonstrate substantial white-box hallucination increases on several models, alongside smaller and more heterogeneous cross-model transfer gains.

Critical strengths
  • Inducing plausible unsupported descriptions is a consequential failure mode that differs from forcing a predefined string or suppressing generation. The formulation in §1 addresses an important aspect of video-language model reliability.
  • The method is coherent and reasonably well specified. Holding the teacher-forced prefix fixed makes the clean-versus-corrupted distribution comparison interpretable. Equations 6–19, Algorithm 1, and the transformation definitions in Appendix A.2 provide a concrete implementation.
  • Several white-box effects are substantial. On Qwen3-VL/VATEX, HSR increases from 31.13 to 80.26, while HUR increases from 2.79 to 8.84 (Table 1). Reporting normalized metrics alongside absolute hallucination counts provides evidence beyond simply generating longer responses.
Critical weaknesses
  • The data split is unclear. Section 4.1 describes two subsets of 100 videos each, but does not clarify whether these evaluation videos were also used for hyperparameter tuning or method development. In addition, the statistical significance of the reported results is unclear.

  • Equations 15–17 simplify to �0��0�. However, Table 3 does not establish whether amplifying this distributional difference improves on simply matching the corrupted-video distribution, �1��1�. The authors should compare these targets while keeping the selected corrupted video, position weights, and optimization procedure unchanged. This would isolate the value of the contrastive construction. Its relationship to visual contrastive decoding also needs clearer discussion.

  • For Qwen3-VL, Quality decreases from 9.16 to 8.30 on VATEX and from 9.89 to 8.52 on MSRVTT (Table 1). The second judge reports larger decreases (Table 8). These outputs may remain readable, but the results indicate a model-dependent tradeoff between hallucination induction and linguistic quality. The claim should be qualified or supported by paired human judgments.

Questions for authors

Please refer to Critical Weaknesses. If I've missed anything, don't hesitate to let me know.

2 comments

Thank you for your rebuttal, I’d like to to raise my initial rating.

Thank you for your recognition and active engagement in the discussion!

Reviewer_cr55OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies adversarial perturbations that induce open-ended hallucinations in Video-LLMs without specifying the desired hallucinated response. VidHaze first generates a clean caption, finds a disruptive spatial-temporal corruption, uses the difference between clean and corrupted token distributions to construct hallucination-oriented token preferences, and then optimizes a small perturbation to the original video toward those preferences. The attack is evaluated on two video datasets and four Video-LLMs using sentence-/unit-level hallucination rates and language-quality metrics under white- and black-box settings.

Critical strengths
  1. The paper identifies an empirical difference in the sensitivity of correct and hallucinated Video-LLM predictions
  2. The paper provides a reasonably extensive empirical evaluation across multiple models and white-/black-box settings.
Critical weaknesses
  • Overall, the paper tries to look fancy by decorating with redundant equations and math notations. For example, the right hand side of Eq. 17 is basically equals to 2∗lth−ltc2*l^h_t - l^c_t.
  • Controlling a small subset of vocab (eg, “a”, “sorry”) is not a standard of practice, especially given that the paper mentioned the open-ended generation.
  • I did not understand why we need to induce hallucinations.
  • There is no human validation or filtering on any phases, while every phase has potential risks of multiple errors introduced. For example, hallucination can happen on video captioning on phase 1.
  • There are no new scientific findings from the experiment.
Questions for authors
  • Why does it have to be Thompson sampling rather than Bernoulli sampling? Can authors compare the impact of the sampling strategy shift?
  • What is the use case of such generated hard perturbations?
5 comments
AuthorsOpenReview ↗

Looking Forward to Your Engagement in the Rebuttal Discussion

Thank you for your initial feedback. We believe we have thoroughly addressed your concerns through additional experiments and clarifications. We would appreciate your continued engagement in the discussion.

I appreciate the authors' comprehensive rebuttal, especially the additional experiments on Bernoulli Sampling, and the refusal subset. The additional questions are presented below. Also, I acknowledge that my first reading was inaccurate. The remaining concern is the novelty and W5.

  • W5: Xuan et al. [1] already provides similar observations on images. In my opinion, it is a subjective matter whether the observation holds on video domain is novel by itself since the observed behavior is expected to generalize to videos.
  • Can authors explain more on Qual. metric? I am not sure how to interpret the table that VidHaze tend to score the lowest among the compared methods.

[1] PAS : Prelim Attention Score for Detecting Object Hallucinations in Large Vision--Language Models

AuthorsOpenReview ↗

Thanks for the reviewer's further feedback.

Q1: Clarifying Similarity to PAS [1]

A1: After reading PAS [1], we believe it does not undermine our novelty, as the two works investigate different phenomena and pursue different methodological objectives. We suspect that the perceived similarity may stem primarily from their motivation experiments (PAS's motivation-stage MI analysis does not carry over to its final attention-based detector, whereas our observed sensitivity gap directly informs our attack design). Specifically, the two motivation experiments share two superficial similarities:

  1. Both analyze next-token probability distributions under a fixed generation prefix.
  2. Both compare the prediction differences between correct and hallucinated content.

However, these superficial similarities do not imply that the two works investigate the same phenomenon. Specifically:

(1) Different visual conditions and research objectives. PAS compares original-image-conditioned predictions with predictions averaged over reference images, whereas we compare predictions for the same video before and after controlled spatiotemporal corruptions under an identical text prefix. PAS studies dependence on image information, whereas we investigate sensitivity to spatiotemporal corruptions.

(2) Findings on images cannot be assumed to generalize to videos without quantitative evidence. Videos introduce temporal dependencies and motion absent in static images. We evaluate four Video-LLMs under three spatiotemporal corruption schemes, observing higher average nJSD for correct than hallucinated sentences across all 12 combinations.

(3) Different mathematical metrics. PAS uses Conditional Mutual Information (MI), defined as DMI(yk)=−I(V;Yk∣y<k,t)D_{\mathrm{MI}}(y_k)=-I(V;Y_k\mid y_{<k},t). Its image-marginalized distribution is approximated by averaging predictions over reference images: p(yk∣y<k,t)≈1∣I∣∑I∈Ip(yk∣y<k,(I,t))p(y_k\mid y_{<k},t)\approx\frac{1}{|\mathcal I|}\sum_{I\in\mathcal I}p(y_k\mid y_{<k},(I,t)).

In contrast, we use normalized Jensen-Shannon Divergence (nJSD), defined as Ji,ja=1∣Ii,j∣log⁡2∑t∈Ii,jJSD⁡(pclean(t),pcorrupta(t))\mathcal J_{i,j}^{a}=\frac{1}{|\mathcal I_{i,j}|\log 2}\sum_{t\in\mathcal I_{i,j}}\operatorname{JSD}\left(p_{\mathrm{clean}}(t),p_{\mathrm{corrupt}}^{a}(t)\right).

The only direct formal similarity is that both use next-token probability distributions under a fixed generation prefix; their metric definitions, distributional comparisons, and the phenomena they characterize are fundamentally different.

(4) Other notable differences include:

  • Task: PAS targets object-level hallucination detection in static images, whereas we study sentence-level open-ended hallucination induction in videos.
  • Method: PAS's final detection score is based on attention weights, rather than the mutual-information analysis used in its motivation study. We directly translate our observed sensitivity gap into hard-negative construction and adversarial optimization, combining Thompson Sampling, distributional contrast, and PGD.
  • Modality: PAS focuses on Image-LLMs, whereas we investigate Video-LLMs with both spatial and temporal dependencies.

Q2: Explaining Qual. Metric

A2: We acknowledge that VidHaze obtains the lowest Qual score. However, Qual is only one metric for assessing language quality, and the marginal Qual difference is negligible compared with VidHaze's gains in hallucination-induction effectiveness. Specifically:

  1. VidHaze achieves the best PPL, providing complementary evidence of its language quality. Qual evaluates holistic paragraph-level fluency, while PPL measures token-level predictability. We report both for a more comprehensive evaluation. Notably, VidHaze achieves the best PPL (15.315) among all compared methods.

  2. The Qual gap between VidHaze and the next-lowest attack baseline is marginal. VidHaze's Qual (9.183) is only 0.042 lower than Mirage (9.225) on a 10-point scale. We encourage the reviewer to apply the prompt in Appendix A.8 to the generated responses and directly compare the resulting scores across responses.

  3. VidHaze achieves stronger hallucination induction with only a marginal Qual trade-off. Averaged across all settings, it outperforms Mirage by 3.64 pp in HSR and 0.38 pp in HUR (+8.14% and +9.96%), demonstrating its ability to maintain comparable language quality while substantially improving attack effectiveness.

MethodQual ↑PPL ↓HSR ↑HUR ↑
Clean9.58716.16625.2301.986
Random9.60816.18026.4082.052
R-Mute-N9.39415.45543.3023.614
Mirage9.22515.59244.6933.783
TVA9.39915.37837.0402.983
VidHaze9.18315.31548.3334.159

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.