Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Deliberately corrupted video features, used as contrast during decoding, reduce hallucinated outputs from frozen video-language models.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 06:31 UTC pith:KY6JLRFO

load-bearing objection The full-text SSCD method is a plausible new idea that deserves scrutiny, but the abstract doesn't match the body, the key mechanism is unvalidated, and the lack of error bars/code makes the empirical claims provisional. the 4 major comments →

arxiv 2601.22574 v2 pith:KY6JLRFO submitted 2026-01-30 cs.CV cs.AI

Enhancing Video Representations with Spatiotemporal-Semantic Residual to Mitigate Hallucinations in Video Large Multimodal Models

Yuansheng Gao , Jinman Zhao , Tong Zhang , Xingguo Xu , Wenbin Xing , Han Bao , Zonghui Wang , Wenzhi Chen This is my paper
classification cs.CV cs.AI
keywords video hallucinationcontrastive decodingspatiotemporal consistencyrandom walkconditional mutual informationvideo language modelresidual disruptioninference-time intervention
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that hallucinations in video large multimodal models can be suppressed without retraining the model, by learning a small residual network that deliberately corrupts the video representation and then using the corrupted version as a negative contrast at decoding time. The corruption is not random: it targets two properties the authors argue drive hallucination—spatiotemporal consistency (a video feature should not be too self-consistent across frames) and semantic alignment with the asked question/answer. On top of a frozen video-LLM, a lightweight 'disruptor' is trained to break both, and the final token distribution is obtained by subtracting the corrupted-model logits from the original logits, subject to a plausibility filter. The paper reports consistent hallucination reduction on VideoHallucer, EventHallusion, and VideoHallu, while largely preserving—and on one backbone even improving—general video QA and reasoning performance. A sympathetic reader would care because this is a parameter-efficient, single-forward-pass intervention that could make existing deployed video chatbots more reliable.

Core claim

The paper's central discovery is that a trained residual perturbation can convert a faithful video representation into a 'hallucination-prone' one, and that this negative representation is a useful control signal: subtracting its logits during decoding removes hallucinated content while keeping grounded content. The authors formalize hallucination-proneness in two ways. First, they break spatiotemporal consistency by minimizing a random-walk cycle-consistency score over a graph of video tokens, deliberately making cross-frame transitions less predictable. Second, they break semantic consistency by minimizing conditional mutual information between the perturbed visual tokens and the ground-tr

What carries the argument

The Spatiotemporal-Semantic Disruptor (SSD) is the load-bearing component: a small MLP that outputs a residual added to frozen video features. Training optimizes a combined loss: (1) a spatiotemporal disruption loss, computed as the average log cycle-consistency score of palindromic random walks on the spatiotemporal graph (Eq. 10-11), which the paper minimizes to weaken long-range self-consistency; and (2) a semantic disruption loss, the negative log-likelihood of the ground-truth answer under the negative features (Eq. 12-15), standing in for conditional mutual information. At inference, the trained disruptor produces negative features in one forward pass; contrastive decoding (Eq. 17-19)

Load-bearing premise

The whole scheme rests on the assumption that the cycle-consistency and conditional-mutual-information objectives produce negative features that are specifically 'hallucination-prone' in a way that matters to the LLM—if they only add arbitrary noise, the contrastive decoding might still appear to work but the explanation would be wrong.

What would settle it

Train the disruptor with the two loss terms replaced by random Gaussian noise or by frame shuffling, while keeping the same contrastive decoding. If hallucination reduction on a held-out benchmark stays at the same level, the learned spatiotemporal and semantic objectives are not the operative cause of the improvement.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Frozen video-LLMs can be made less hallucination-prone by adding a small trainable residual disruptor and switching to contrastive decoding, without any modification to the backbone or its training data.
  • Because the negative features come from the same forward pass (a single extra lightweight MLP), the method avoids the multi-sample or extra-forward-pass latency of prior contrastive-decoding interventions.
  • On the two backbones tested, the method either improves or roughly preserves performance on general QA and reasoning benchmarks, suggesting hallucination reduction need not come at the cost of overall capability.
  • The two objectives are complementary rather than redundant: ablations show spatiotemporal disruption alone gives a clear gain, and adding semantic disruption yields a further improvement on event-level hallucination benchmarks.
  • If the characterization is right, 'hallucination-prone features' are not a pathology of the LLM alone but a manipulable property of the video representation, opening a new axis for diagnosis and control.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper does not prove that low cycle-consistency specifically corresponds to hallucination; a testable extension would compare the learned disruptor against simpler corruptions (frame shuffling, Gaussian noise, feature dropout) to see whether the random-walk objective is doing the work or whether any strong perturbation suffices.
  • The semantic disruption term requires ground-truth answers during training; a natural extension is to train with only the question or with self-generated answers, and check whether the hallucination-mitigation effect survives, which would determine how much supervision the method really needs.
  • The contrastive strength is tuned separately per backbone and the paper notes sensitivity to it; an adaptive schedule that scales the strength by, say, the model's uncertainty or the divergence between the two distributions could make the method more portable across architectures and tasks.
  • The method's success raises the possibility that similar residual-disruption contrastive decoding could be applied to other failure modes (e.g., object hallucination in images, or temporal ordering errors) by defining the right consistency objective for that modality.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes SSCD (also called ViSSRes in the abstract), a decoding-time intervention for Video Large Multimodal Models. A lightweight residual disruptor is trained on ShareGPT4Video captions to produce negative video features by (i) minimizing a random-walk cycle-consistency score that is intended to break spatiotemporal coherence and (ii) minimizing the conditional mutual information / log-likelihood of the correct answer under the negative features to weaken semantic alignment. At inference, the logits from these negative features are subtracted from the original logits under an adaptive plausibility constraint (Eqs. 17–19). Experiments on VideoHallucer, EventHallusion, VideoHallu, ActivityNet-QA, and MMVU compare against TCD, MotionCD, and Dino-Heal, reporting overall improvements on hallucination benchmarks and mixed results on general understanding benchmarks.

Significance. If the mechanism is valid, the approach offers a promising lightweight, frozen-backbone intervention for reducing video hallucination without retraining the VLM. The paper includes evaluations on multiple standard benchmarks and ablates the two disruption objectives. A strength is that the disruptor is trained on external captions and tested on separate benchmarks, so the design is not circular in the strong sense. However, the core hypothesis that disrupting spatiotemporal cycle consistency produces features that induce hallucinations is not empirically validated, and the reported results are weakened by hyperparameter selection on the same benchmarks used for final evaluation and by the absence of statistical confidence measures.

major comments (4)
  1. [§4.1, Eq. 10–11] The central premise is that low cycle-consistency in the random-walk objective corresponds to hallucination-prone video features, but no theoretical or empirical evidence supports this. The paper never shows that minimizing the palindromic walk return probability degrades visual information in a way that induces hallucinations, as opposed to simply adding noise. The ablation in Table 4 is ambiguous: on LLaVA-NeXT, adding S1 alone improves Overall by only 2.2 points (48.7→50.9) and actually lowers Entire (46.5→44.7). A control condition with generic feature corruption (e.g., Gaussian noise or frame shuffling) is needed to attribute the gain to spatiotemporal disruption rather than residual perturbation.
  2. [§5.1 and Figures 2–3] Hyperparameters λ, α, and β are selected through ablations on MMVU (Direct subset) and EventHallusion, which are exactly the benchmarks reported in the final results (Tables 1 and 3). This constitutes tuning on the evaluation sets and can inflate the reported gains. The authors should either use a separate validation set (e.g., a held-out split of the training data or a different benchmark) or demonstrate that the results are stable across a range of hyperparameter values without choosing the best on the evaluation benchmarks. At minimum, the hyperparameter selection procedure must be disclosed as a limitation.
  3. [Abstract vs. full text] The abstract claims that the method 'reduces the hallucination rate of LLaVA-NeXT-Video on EventHallusion by 40.69% and improves video understanding on MMVU by 18.36% under the CoT setting.' These numbers do not appear anywhere in the full text. Table 1 shows EventHallusion Overall for LLaVA-NeXT-Video moving from 48.7 to 51.3 (an absolute improvement of 2.6 points, or roughly 5% relative on the hallucination rate), and Table 3 shows MMVU CoT moving from 30.5 to 30.1 (a decrease). The abstract also claims 'a single forward pass' at inference, but Eq. 17–19 explicitly require evaluating the LLM on both the original and negative visual features, i.e., two forward passes. These discrepancies must be corrected or thoroughly explained.
  4. [§5.2, Tables 1–3] No error bars, confidence intervals, or significance tests are reported. Many differences are small (e.g., Table 3: LLaVA-NeXT ActivityNet-QA Accuracy 53.3 vs. 53.3; MMVU Direct 31.3 vs. 31.1). Because the disruptor training involves stochastic optimization, results should be reported over multiple random seeds (at least 3) with mean and standard deviation. Without this, it is not possible to assess whether the observed improvements are reliable or within noise.
minor comments (5)
  1. [General] The method name is inconsistent: the abstract and the provided arXiv title use 'ViSSRes', while the full manuscript uses 'SSCD'. Please harmonize the terminology and ensure the title matches the manuscript content.
  2. [Figure 2] The caption and y-axis label refer to 'Overall Accuracy', but the plot is for the direct multiple-choice subset of MMVU; clarify the exact metric and setting.
  3. [§2.1] The citation 'Gao et al., 2025b' (the first author's own work on mental manipulation detection) seems out of place in the related work on video LLMs; consider replacing it with a more representative reference for LLM backbones.
  4. [§4.2, Eq. 15] The semantic disruption loss minimizes the log-likelihood of the correct answer. If the negative features become very poor, the probability can saturate near zero, potentially causing numerical instability or vanishing gradients. Please discuss or add a small floor.
  5. [Table 1] There are formatting issues, such as missing spaces in several rows (e.g., '9.526.0' and '29.859.644.1'). Please correct the typesetting.

Circularity Check

0 steps flagged

No circular derivation; only an incidental non-load-bearing self-citation and validation caveats.

full rationale

The derivation chain is not circular. The disruptor M is trained on 3,000 random ShareGPT4Video instances with two objectives: the spatiotemporal cycle-consistency loss L_T (Eq. 11) and the semantic negative-log-likelihood loss L_S (Eq. 15), while the LLM backbone and projector are frozen. Evaluation is performed on disjoint external benchmarks (VideoHallucer, EventHallusion, VideoHallu, ActivityNet-QA, MMVU); no benchmark label is used to fit the disruptor. Eq. 15 trains negative features to reduce p_theta(Y | Z_v^neg) for the ground-truth caption, and Eq. 17 subtracts those negative logits during decoding; this is the designed mechanism of contrastive decoding, not a tautology, because the reported hallucination reductions are empirical outcomes that could have failed. The weakest point is Section 4.1's assertion that low cycle-consistency 'steer[s] them toward hallucination-prone states during decoding'—this is an unvalidated proxy assumption and a correctness risk, but not a circular reduction, since no equation forces the benchmark improvement. A further validation concern is that alpha and beta are ablated on EventHallusion (Figure 3) and then used in the Table 1 results, which is test-set hyperparameter selection rather than constructional circularity. The only self-citation (Gao et al., 2025b in Section 2.1) is an incidental reference in a general list of LLM works and is not load-bearing. There is no imported uniqueness theorem, no ansatz smuggled through self-citation, and no renaming of a known result as a derivation. The abstract's claimed 40.69% and 18.36% improvements do not obviously match Tables 1 and 3, but that is a reporting/correctness issue, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central claim rests on two heuristic mappings: cycle-consistency loss to hallucination-proneness, and frozen-LLM log-likelihood to conditional mutual information. In addition, four hyperparameters are selected on the evaluation benchmarks themselves, so the reported numbers partly reflect tuning rather than prediction.

free parameters (5)
  • τ (transition temperature, Eq. 7) = 0.07
    Controls the sharpness of the random-walk transition matrix; taken from Jabri et al. (2020) rather than derived.
  • λ (loss weight, Eq. 16) = 5
    Balances spatiotemporal and semantic disruption losses; selected via ablation on the MMVU direct multiple-choice subset, which is also a main evaluation benchmark.
  • α (contrastive strength, Eq. 17) = 0.8 for Video-LLaVA, 0.4 for LLaVA-NeXT
    Controls how strongly negative logits are subtracted; selected via ablation on EventHallusion, which is also a main evaluation benchmark.
  • β (plausibility truncation, Eq. 18) = 0.1
    Controls candidate token truncation in contrastive decoding; selected via ablation on EventHallusion.
  • learning rate = 1e-6 (Video-LLaVA), 5e-6 (LLaVA-NeXT)
    Per-backbone learning rate for disruptor training; another tuned hyperparameter.
axioms (5)
  • ad hoc to paper Disrupting spatiotemporal cycle consistency of video features yields negative features that induce hallucinations.
    The random-walk loss LT (Eq. 11) only minimizes a consistency score; the paper asserts, without proof or an isolated probe, that this makes the LLM hallucinate.
  • domain assumption Minimizing the frozen LLM's log-likelihood of the ground-truth answer under negative features approximates minimizing conditional mutual information.
    Eqs. 12–15 replace a true conditional entropy with model log-likelihood; the text also drops a minus sign when calling log-likelihood 'conditional entropy'.
  • domain assumption Contrastive decoding with a plausibility constraint suppresses hallucinated tokens without degrading general abilities.
    Adopted from prior contrastive decoding work (Leng et al., 2024; Chen et al., 2025); not derived here.
  • domain assumption A Markov random walk over frame-level token affinities faithfully models the spatiotemporal consistency that matters for hallucination.
    The graph construction in Eqs. 6–8 is mathematically standard, but its adequacy as a model of video semantics is assumed.
  • domain assumption A disruptor trained to lower the likelihood of training captions will disrupt semantic alignment for unseen questions.
    The semantic loss LS (Eq. 15) requires ground-truth answers during training; at inference no answer is available, so transfer is assumed.
invented entities (2)
  • Spatiotemporal-Semantic Disruptor (M) no independent evidence
    purpose: A light MLP that adds a residual perturbation to frozen video features to produce negative features H_v^neg (Eq. 5).
    No prediction outside the paper's benchmarks; its behavior is defined entirely by the training losses.
  • Negative video features H_v^neg no independent evidence
    purpose: Hallucination-prone video representations used as the contrastive branch in Eq. 17.
    Constructed entity with no independent falsifiable handle outside the paper.

pith-pipeline@v1.3.0-alltime-deepseek · 17119 in / 20134 out tokens · 180123 ms · 2026-08-03T06:31:18.305071+00:00 · methodology

0 comments
read the original abstract

Although Video Large Multimodal Models have achieved strong performance in video understanding, they still suffer from hallucination. Existing inference-time intervention methods usually modify videos under the contrastive decoding framework, but their heuristic designs bring limited improvements and increase inference latency. To address these issues, we propose ViSSRes, an inference-time intervention method that enhances video representations through a lightweight MLP-style network. Specifically, we use a contrastive random walk approach to characterize the spatiotemporal consistency of video representations, and introduce conditional mutual information to associate video representations with the model's semantic understanding. With the model backbone kept frozen, ViSSRes learns residuals for video representations and optimizes them from both spatiotemporal and semantic consistency perspectives. During inference, ViSSRes requires only a single forward pass and introduces no substantial additional inference cost. Experiments show that ViSSRes reduces the hallucination rate of LLaVA-NeXT-Video on EventHallusion by 40.69% and improves video understanding on MMVU by 18.36% under the CoT setting, demonstrating its effectiveness in mitigating hallucinations.

Figures

Figures reproduced from arXiv: 2601.22574 by Han Bao, Jinman Zhao, Tong Zhang, Wenbin Xing, Wenzhi Chen, Xingguo Xu, Yuansheng Gao, Zonghui Wang.

Figure 1
Figure 1. Figure 1: Overview of the proposed SSCD. Left: We freeze the VideoLLM and train only a lightweight spatiotemporal-semantic disruptor. Right: During inference, we use the trained disruptor to generate negative video features with disrupted spatiotemporal and semantic consistency, and mitigate hallucinations via calibrated distribution sampling with contrastive decoding. the resulting 2z-step round-trip transition fro… view at source ↗
Figure 2
Figure 2. Figure 2: Ablation analysis of λ in training on the direct multiple￾choice subset of MMVU with Video-LLaVA as the backbone. 50.1 51.6 51.1 53.8 52.6 52.6 50.0 51.0 52.0 53.0 54.0 0.5 0.6 0.7 0.8 0.9 1.0 Overall Accuracy (%) Value of α 53.5 53.3 52.8 53.8 53.1 52.8 50.0 51.0 52.0 53.0 54.0 0 0.001 0.01 0.1 0.2 0.5 Overall Accuracy (%) Value of β [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation analysis of α and β in contrastive decoding on EventHallusion using Video-LLaVA as the backbone. LLaVA. Since the parameters τ in Eq. 7 and λ in Eq. 16 only influence the gradients of the loss terms during train￾ing, we fix τ and vary λ to examine its impact on the direct multiple-choice subset of MMVU. As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: An example from the VideoHallu dataset with LLaVA￾NeXT-Video as the backbone, where SSCD demonstrates superior temporal information modeling compared to the baseline. predictions of SSCD are compared against those of several baseline methods. In the first example shown in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Overview of the architecture of SSD. B. Evaluation and Benchmarks VideoHallucer Evaluation VideoHallucer comprises both basic questions and hallucination-oriented questions, enabling not only the assessment of hallucination severity but also the analysis of model biases. Specifically, VideoHallucer evaluates multiple types of hallucinations, including Object-Relation Hallucination (ORH), Temporal Hallucina… view at source ↗
Figure 7
Figure 7. Figure 7: An example from the EventHallusion (left) and VideoHallucer (right) datasets using LLaVA-NeXT-Video as the backbone. Question: Do any dogs show inconsistent body or leg proportions throughout the video? Ground-truth: No. SSCD: No, there are no inconsistent body or leg proportions shown by any of the dogs in the video. Vanilla, TCD, and Dino-Heal: Yes, one of the dogs has a very short tail throughout the vi… view at source ↗
Figure 8
Figure 8. Figure 8: An example from the VideoHallu dataset using LLaVA-NeXT-Video as the backbone. Question: Is the athlete in the video indoors? Ground-truth: No. Vanilla, TCD, and Dino-Heal: Yes, the athlete is seen running down a track and jumping over a bar indoors. MotionCD: Yes, the video shows that the athlete is jumping over a bar indoors. SSC D : No, the athlete is not indoors. The video shows him running and jumping… view at source ↗
Figure 9
Figure 9. Figure 9: An example from the ActivityNet-QA dataset using Video-LLaVA as the backbone. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: An example from the MMVU dataset using Video-LLaVA as the backbone. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Tracking the Truth: Object-Centric Spatio-Temporal Monitoring for Video Large Language Models

    cs.CV 2026-05 unverdicted novelty 7.0

    STEMO-Bench evaluates intermediate spatio-temporal reasoning in video MLLMs via object-centric facts, and STEMO-Track improves consistency by chunk-wise trajectory construction and aggregation.

Reference graph

Works this paper leans on

22 extracted references · 7 linked inside Pith · cited by 1 Pith paper

  1. [4]

    Grounding language with vision: A conditional mutual information calibrated decoding strat- egy for reducing hallucinations in lvlms.arXiv preprint arXiv:2505.19678,

    Fang, H., Zhou, C., Kong, J., Gao, K., Chen, B., Liang, T., Ma, G., and Xia, S.-T. Grounding language with vision: A conditional mutual information calibrated decoding strat- egy for reducing hallucinations in lvlms.arXiv preprint arXiv:2505.19678,

    arXiv

  2. [5]

    Exploring hallucination of large multimodal models in video understanding: Benchmark, analysis and mitigation.arXiv preprint arXiv:2503.19622, 2025a

    Gao, H., Qu, J., Tang, J., Bi, B., Liu, Y ., Chen, H., Liang, L., Su, L., and Huang, Q. Exploring hallucination of large multimodal models in video understanding: Benchmark, analysis and mitigation.arXiv preprint arXiv:2503.19622, 2025a. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al. Llama-adapter v2: ...

    Pith/arXiv arXiv

  3. [6]

    Mentalmac: Enhancing large language models for detect- ing mental manipulation via multi-task anti-curriculum distillation.arXiv preprint arXiv:2505.15255, 2025b

    Gao, Y ., Bao, H., Zhang, T., Li, B., Wang, Z., and Chen, W. Mentalmac: Enhancing large language models for detect- ing mental manipulation via multi-task anti-curriculum distillation.arXiv preprint arXiv:2505.15255, 2025b. Gunjal, A., Yin, J., and Bas, E. Detecting and prevent- ing hallucinations in large vision language models. In Proceedings of the AAA...

    arXiv

  4. [7]

    Vistadpo: Video hierarchical spatial-temporal direct preference optimization for large video models.arXiv preprint arXiv:2504.13122,

    Huang, H., Chen, H., Wu, S., Luo, M., Fu, J., Du, X., Zhang, H., and Fei, H. Vistadpo: Video hierarchical spatial-temporal direct preference optimization for large video models.arXiv preprint arXiv:2504.13122,

    Pith/arXiv arXiv

  5. [8]

    H., Jo, Y ., and Seo, M

    Lee, S., Park, S. H., Jo, Y ., and Seo, M. V olcano: Mitigating multimodal hallucination through self-feedback guided revision. In2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL 2024, pp. 391–

    2024

  6. [10]

    Mitigating hallucination in large multi-modal models via robust instruction tuning

    Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y ., and Wang, L. Mitigating hallucination in large multi-modal models via robust instruction tuning. InThe Twelfth International Conference on Learning Representations, 2024a. Liu, Y ., Li, S., Liu, Y ., Wang, Y ., Ren, S., Li, L., Chen, S., Sun, X., and Hou, L. Tempcompass: Do video llms really understand vide...

    Pith/arXiv arXiv 2024

  7. [11]

    Countervid: Counterfactual video generation for mitigat- ing action and temporal hallucinations in video-language models.arXiv preprint arXiv:2601.04778,

    Poppi, T., Uzkent, B., Garg, A., Porto, L., Kessler, G., Yang, Y ., Cornia, M., Baraldi, L., Cucchiara, R., and Schiffers, F. Countervid: Counterfactual video generation for mitigat- ing action and temporal hallucinations in video-language models.arXiv preprint arXiv:2601.04778,

    arXiv

  8. [12]

    Smart- sight: Mitigating hallucination in video-llms without com- promising video understanding via temporal attention collapse.arXiv preprint arXiv:2512.18671,

    Sun, Y ., Zhang, M., Li, F., Hong, G., and Yang, M. Smart- sight: Mitigating hallucination in video-llms without com- promising video understanding via temporal attention collapse.arXiv preprint arXiv:2512.18671,

    arXiv

  9. [14]

    Videohallucer: Evaluating intrinsic and extrinsic halluci- nations in large video-language models.arXiv preprint arXiv:2406.16338,

    Wang, Y ., Wang, Y ., Zhao, D., Xie, C., and Zheng, Z. Videohallucer: Evaluating intrinsic and extrinsic halluci- nations in large video-language models.arXiv preprint arXiv:2406.16338,

    Pith/arXiv arXiv

  10. [15]

    Wu, C.-H., Chang, K.-P., Sheng, Y .-Y ., Chung, H.-K., Wang, K.-C., and Wang, Y .-C. F. Season: Mitigating temporal hallucination in video large language models via self-diagnostic contrastive decoding.arXiv preprint arXiv:2512.04643, 2025a. 10 Mitigating Hallucinations in Video Large Language Models via Spatiotemporal-Semantic Contrastive Decoding Wu, Z....

    arXiv

  11. [17]

    Kardia-r1: Unleashing llms to reason to- ward understanding and empathy for emotional support via rubric-as-judge reinforcement learning.arXiv preprint arXiv:2512.01282, 2025a

    Yuan, J., Cui, Z., Wang, H., Gao, Y ., Zhou, Y ., and Naseem, U. Kardia-r1: Unleashing llms to reason to- ward understanding and empathy for emotional support via rubric-as-judge reinforcement learning.arXiv preprint arXiv:2512.01282, 2025a. Yuan, J., Du, D., Zhang, H., Di, Z., and Naseem, U. Re- versal of thought: Enhancing large language models with pre...

    arXiv

  12. [18]

    Video-llama: An instruction- tuned audio-visual language model for video understand- ing

    Zhang, H., Li, X., and Bing, L. Video-llama: An instruction- tuned audio-visual language model for video understand- ing. InProceedings of the 2023 Conference on Empir- ical Methods in Natural Language Processing: System Demonstrations, pp. 543–553,

    2023

  13. [19]

    Eventhallusion: Diagnosing event hallucinations in video llms.arXiv preprint arXiv:2409.16597, 2024a

    Zhang, J., Jiao, Y ., Chen, S., Zhao, N., Tan, Z., Li, H., and Chen, J. Eventhallusion: Diagnosing event hallucinations in video llms.arXiv preprint arXiv:2409.16597, 2024a. Zhang, S., Fang, Q., Yang, Z., and Feng, Y . Llava-mini: Effi- cient image and video large multimodal models with one vision token. InThe Thirteenth International Conference on Learni...

    arXiv

  14. [20]

    j., Gui, L., Fu, D., Feng, J., Liu, Z., and Li, C

    Zhang, Y ., Li, B., Liu, h., Lee, Y . j., Gui, L., Fu, D., Feng, J., Liu, Z., and Li, C. Llava-next: A strong zero-shot video understanding model, April 2024b. URL https://llava-vl.github.io/ blog/2024-04-30-llava-next-video/. Zhao, S., Wang, Y ., Yang, Z., and Cai, D. Region mutual information loss for semantic segmentation.Advances in Neural Information...

    2024

  15. [21]

    Can pruning improve reasoning? revisiting long-cot compression with capability in mind for better reasoning

    Zhao, S., Yuan, J., Wu, J., Wang, Z., Yang, G., and Naseem, U. Can pruning improve reasoning? revisiting long-cot compression with capability in mind for better reasoning. arXiv preprint arXiv:2505.14582, 2025a. Zhao, Y ., Zhang, H., Xie, L., Hu, T., Gan, G., Long, Y ., Hu, Z., Chen, W., Li, C., Xu, Z., Wang, C., Shangguan, Z., Liang, Z., Liu, Y ., Zhao, ...

    arXiv

  16. [22]

    Layernorm Linear GeLU Linear GeLU Linear Tanh Layernorm Linear GeLU Linear GeLU Linear Tanh Figure 6.Overview of the architecture of SSD

    Notably, this learnable component constitutes only a small portion of the VideoLLM’s total parameters, resulting in minimal computational overhead. Layernorm Linear GeLU Linear GeLU Linear Tanh Layernorm Linear GeLU Linear GeLU Linear Tanh Figure 6.Overview of the architecture of SSD. B. Evaluation and Benchmarks VideoHallucer EvaluationVideoHallucer comp...

    2024

  17. [2017]

    Affordance- r1: Reinforcement learning for generalizable affordance reasoning in multimodal large language model.arXiv preprint arXiv:2508.06206,

    Wang, H., Wang, S., Zhong, Y ., Yang, Z., Wang, J., Cui, Z., Yuan, J., Han, Y ., Liu, M., and Ma, Y . Affordance- r1: Reinforcement learning for generalizable affordance reasoning in multimodal large language model.arXiv preprint arXiv:2508.06206,

    Pith/arXiv arXiv

  18. [2019]

    Helpd: Mitigating hallucination of lvlms by hierarchical feedback learning with vision-enhanced penalty decoding

    Yuan, F., Qin, C., Xu, X., and Li, P. Helpd: Mitigating hallucination of lvlms by hierarchical feedback learning with vision-enhanced penalty decoding. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pp. 1768–1785,

    2024

  19. [2022]

    Li, Z., Shi, H., Gao, Y ., Liu, D., Wang, Z., Chen, Y ., Liu, T., Zhao, L., Wang, H., and Metaxas, D. N. The hidden life of tokens: Reducing hallucination of large vision-language models via visual information steering. InForty-second International Conference on Machine Learning, 2025b. Li, Z., Wu, X., Shi, G., Qin, Y ., Du, H., Liu, F., Zhou, T., Manocha...

    2024

  20. [2023]

    Mitigating hallucination in videollms via temporal-aware activation engineering.arXiv preprint arXiv:2505.12826,

    Cai, J., Zhou, W., Zhang, Z., Hong, J., Zhan, N., and Li, H. Mitigating hallucination in videollms via temporal-aware activation engineering.arXiv preprint arXiv:2505.12826,

    Pith/arXiv arXiv

  21. [2024]

    Decoupling contrastive decoding: Robust hallucination mitigation in multimodal large language models.arXiv preprint arXiv:2504.08809,

    Chen, W., Yan, X., Wen, B., Yang, F., Gao, T., Zhang, D., and Chen, L. Decoupling contrastive decoding: Robust hallucination mitigation in multimodal large language models.arXiv preprint arXiv:2504.08809,

    arXiv

  22. [2025]

    Pami-vdpo: Mitigating video hallucinations by prompt- aware multi-instance video preference learning.arXiv preprint arXiv:2504.05810,

    Ding, X., Zhang, K., Han, J., Hong, L., Xu, H., and Li, X. Pami-vdpo: Mitigating video hallucinations by prompt- aware multi-instance video preference learning.arXiv preprint arXiv:2504.05810,

    Pith/arXiv arXiv