acceptodds
Withdrawn from ICLR 2027

Mechanistic Signs of Modality Collapse in Multimodal Event Understanding

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Modality collapse occurs when a multimodal model relies on one dominant modality and ignores others. So far, this phenomenon has been established behaviorally, by modifying the input prompt and analyzing the model's predictions. Such analyses cannot show *whether* a model's computational flow actually leveraged information in a modality. In this work, we focus on the task of identifying important moments in football games and mechanistically verify modality collapse through causal intervention. We separately ablate a moment's video and its commentary transcription from the model's internal representations rather than the prompt, and quantify the impact each modality has on the final prediction. Using the signs of these impacts, we place each moment in one of four diagnostic quadrants: synergy, video collapse, commentary collapse, or conflict. Collapse takes opposite directions across the two classes: where it occurs, important moments are decided by the video and non-important ones by the commentary. Even when a model classifies a moment correctly, collapse costs it confidence. Synergy occurs predominantly for important moments and conflict for non-important ones. Within the important class, synergy monotonically decays from prototypically important moments (goals) to contextually important ones (shots-on-target). To complement these intervention-based results, we also observe how each modality's attention share and the model's decision evolve across layers. Together, our findings indicate that collapse is as much a characteristic of the moment as of the model. Future work can build on these insights to causally localize *where* collapse occurs for enabling inference-time mitigation.

Primary area:
interpretability and explainable AI
Keywords:
multimodality, MLLMs, mechanistic interpretability, evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_WaiVOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper builds upon “modality collapse”, a phenomenon in which a multimodal model over-relies on either the visual or textual modality, and mechanistically verifies it by looking into the internal representations of multimodal models when classifying the importance of moments in football matches. The authors apply attention knockout to the Qwen2.5-VL-7B-Instruct and Qwen2.5-Omni-7B models, blocking attention to the video or commentary spans across all layers and attention heads, and quantify each modality’s causal influence by measuring the resulting change in the logit margin between the ground-truth and alternative class. The authors classify each moment into one of four categories: synergy, commentary collapse, video collapse, or conflict, and report that the dominant form of collapse depends on moment importance, as commentary collapse occurs for important moments, whereas video collapse occurs for non-important moments. Collapse also reduces model confidence, even when the prediction is correct, while synergy decreases in strength from goals to corners to shots on target. Finally, analyses of attention share and the logit lens reveal that attention allocation remains similar across moment classes, despite substantial differences in the modalities’ causal influence. The authors conclude that Qwen2.5-VL integrates visual and textual information more effectively than Qwen2.5-Omni.

Critical strengths

The paper operationalizes modality use by assigning each moment to a category based on the signs of the measured modality effects, a simplification that makes a qualitative concept measurable and enables systematic comparisons across moments, categories, and layers.

Critical weaknesses
  • The paper never checks whether an underlying class bias could be driving the collapse-direction result, and its collapse label provides no way to distinguish a modality with negligible influence from one that is actively, even if misleadingly, driving the model’s answer.

  • Impact is measured as the change in Zy − Zy¯, so its direction is defined relative to the correct label. If a modality consistently pushes predictions toward the “important” class, its impact will appear positive for important moments but negative for non-important moments, producing the opposite-direction pattern across classes that the paper reports as its main finding, even if the modality’s underlying influence is consistent across both classes.

  • The study is limited to binary classification of football moments using two models that share the same language backbone. Without evaluation across additional tasks, domains, or model architectures, it is unclear whether the findings reflect general patterns of modality integration or are specific to this experimental setting.

Questions for authors
  • Can you report modality impact using a class-independent measure, broken down by class? This would help distinguish class-dependent modality use from a consistent modality bias whose measured impact changes with the ground-truth label.
  • What is the model’s accuracy under a prompt-only baseline? How often does the model default to “important” regardless of input? This would clarify how much of the reported class-dependent pattern reflects a default answer prior.
  • In the attention knockout, is the −∞ mask applied to every query position t, for all edges (t, s), or only to t? How does this differ from deleting span S while keeping position IDs fixed?
  • Does Qwen2.5-Omni-7B receive the video’s audio track? Spoken commentary may already be present in the video input before ablation, potentially confusing the comparison between VL and Omni.
1 comment

Many thanks for the careful review. Here are our responses to the points raised under weaknesses:

  1. Two concerns:
    • Class bias. Indeed, we did not check for this in the paper. Using the class-independent measure you suggest (see our answer to Q1), we find that for both classes the video pushes both models towards 'important', while the commentary pushes VL towards 'important' and Omni slightly towards 'non-important'. So, the opposite-direction pattern across the two classes is a consequence of computing the impact relative to the ground-truth label, and we will revise this claim/statement accordingly.
    • Negligible vs. misleading influence. Like we define in L180-181, collapse merges the two by design, since in both cases the modality does not contribute towards the correct prediction. However, most of the collapse moments are misleading: the non-positive impact exceeds the largest impact of our random-span control (Table 3 in Appendix B) for 43 of 57 moments using VL and 60 of 76 using Omni. We will distinguish the two cases in the revision.
  2. We agree, this is exactly the case for the video (see point 1 and our answer to Q1).
  3. We agree that our findings are limited to this setting, and we do not claim that they generalize beyond it. We focus on these two models because the behavioral claims in Surikuchi et al. (2026) were made about them. However, our framework can be adapted to other tasks, domains, and architectures, which we intend to explore in future work.

Here are our responses to the questions:

  1. Thank you for this suggestion. We compute the impact as the change in Z_YES - Z_NO under knockout, so that a positive value means the modality pushes the model towards 'important', irrespective of the ground-truth label. For each class, we report the mean impact and the share of moments pushed towards 'important', and the p-value of a permutation test for the difference between the classes:

    ModelModalityimportant (#103)non-important (#43)p
    VLvideo+2.34 (93%)+1.10 (79%)< 0.001
    VLcommentary+0.98 (68%)+0.72 (65%)0.45
    Omnivideo+1.66 (94%)+1.36 (95%)0.13
    Omnicommentary-0.40 (49%)-0.80 (37%)0.25

    The video pushes both models towards 'important' for both classes, and the push of the commentary does not differ significantly between the classes for either model. The only exception is VL, where the video pushes less strongly towards 'important' for non-important moments.

  2. A prompt-only input is the same for all moments, so it would yield a single answer per model rather than an accuracy. Also, removing both modalities from the prompt shifts the input distribution, which is what our approach avoids (L048-052). The equivalent in our setup is knocking out both spans at once, which we mention under limitations (L453-455) and intend to explore. About the default: with the full input, both models answer YES for 86% of the moments (VL: 95% important, 63% non-important; Omni: 97%, 60%). Knocking out the video lowers these to 77%, 40% (VL) and 83%, 35% (Omni), so the video drives part of this default.

  3. Yes, the mask is applied to every query position t, i.e., to all edges {(t, s) : s ∉\notin S, s ≠\neq t} at every layer and head, like we mention in L150-151. Equations 2 and 4 use the final position t* only for readability. So, no position outside S can read from S, which makes it equivalent to deleting S while keeping the position IDs fixed. We use the mask because it does this without modifying the input prompt (L167-169).

  4. No, in our experiments, Omni does not receive the video's audio track. Both models receive the commentary only as a text transcription, so knocking out the commentary span removes it entirely, and the comparison between VL and Omni is not affected.

Reviewer_DozUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper investigates the modality collapse phenomenon in MLLMs by applying attention knockout respectively to vision and language modalities across all transformer layers. After the knockout, based on the sign changes in logits of the two modalities, the authors categorize the impact as synergy (++), video collapse (-+), commentary collapse (+-), and conflict (--). Experiments are conducted on Qwen2.5-VL-7B-Instruct and Qwen2.5-Omni-7B for classifying important moments in football videos with commentary. Results show that language (commentary) collapses for important moments and vision (video) collapses for non-important ones. In Section 5, the authors provide additional analysis on layer-wise impacts.

Critical strengths

The paper provides empirical results of attention knockout across all transformer layers of vision and language modalities respectively, on Qwen2.5-VL-7B-Instruct and Qwen2.5-Omni-7B, for classifying important moments in football videos with commentary.

Critical weaknesses
  • The novelty of the paper seems unclear besides providing empirical results for the two Qwen models and the football video setup. Since applying attention knockout across all layers or to selected layers of MLLMs to study information flow is already established ([1], [2], [3]), the paper does not differentiate its contribution from these already established work through comparison.

  • On the methodological side, while impact_s measures the effect of an intervention on the output, a negative impact sign does not necessarily mean that the information is not used. It could also mean that the information is used to to push the prediction towards the wrong class. Therefore, more evidence is needed to support similar claims made in the paper.

[1] Zhi Zhang, Srishti Yadav, Fengze Han, and Ekaterina Shutova. Cross-modal information flow in multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 19781–19791, 06 2025.
[2] Omri Kaduri, Shai Bagon, and Tali Dekel. What’s in the image? a deep-dive into the vision of vision language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 14549–14558, 06 2025.
[3] Zhiyu Guo, Hidetaka Kamigaito, and Taro Watanabe. Attention score is not all you need for token importance indicator in KV cache reduction: Value also matters. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (eds.), Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pp. 21158–21166, Miami, Florida, USA, November 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024.emnlp-main.1178. URL https://aclanthology.org/2024.emnlp-main.1178/.

Questions for authors
  • How do you distinguish your research contribution from existing works ([1], [2], [3]) that you have cited in the paper, besides the empirical results on the two Qwen models for football videos?
1 comment

Many thanks for the review. Here are our responses to the points raised under weaknesses:

  1. We do not claim attention knockout itself as our contribution. Like we mention in L093-097, we adopt it from [1] and [2], but for a different purpose. [1] and [2] use it to localize where in the model a modality is integrated or consumed, and report aggregate effects over a dataset. Instead, we use it to test whether each modality is used at all, for each moment: we knock out the video and the commentary independently and assign the moment to one of four quadrants (synergy, commentary collapse, video collapse, conflict) based on the signs of the two impacts. Since the two spans differ in width by ~250×, we also read the impacts against a width-matched control (Appendix B). To our knowledge, this per-moment diagnostic of modality use is not covered by prior work. About [3]: we cite it only for its point that attention weights alone do not indicate a token's influence (L435-436), which our results confirm for both models. We will make this comparison more explicit in the revision.
  2. Like we define in L180-181, collapse covers both cases by design, since in both the modality does not contribute towards the correct prediction. However, we agree that the two should be separated. Using our width-matched control as a reference for how much a knockout changes the prediction by chance, the modality actively pushes the prediction towards the wrong class in 43 of 57 collapse moments using VL and 60 of 76 using Omni, and is negligible in the rest. We will distinguish the two cases in the revision.
Reviewer_4C6fOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary
  • this paper studies modality collapse in multimodal models using attention knockout on a binary classification task of detecting important vs. non-important moments in videos.
  • the idea behind attention knockout is setting the pre-softmax scores to -inf across all decoder layers and query heads corresponding to the video or commentary tokens. setting pre-softmax scores to -inf transforms the attention weights to zero, thereby completely blocking information flow, after which the change in prediction is measured. the change in prediction is categorized into four states: synergy (when both video and commentary impacts are positive), commentary collapse (when video impact is positive but commentary is not), video collapse (when commentary impact is positive but video is not), and conflict (when both impacts are negative / lower the prediction).
Critical strengths
  • broadly, modality collapse is an interesting area of research, and there is clear evidence in the literature that multimodal models often exhibit bias toward one modality over another. exploring research in this direction is therefore of general interest to ICLR.
Critical weaknesses
  • l.041: the definition of modality collapse is somewhat inaccurate. the claim that "multimodal input does not lead to any increase in confidence ... called modality collapse" is not true; it could simply be that the audio modality lacks relevant cues to answer a visual-centric question. for instance, when asking "what is the color of an object?", the color typically has no relationship to the associated audio in standard scenarios, hence not helping with the prediction.

  • Qwen2.5-VL is fairly dated by now. it would be valuable to evaluate newer omni-modal multimodal llms, as more recent models exhibit stronger cross-modal alignment, which is directly tied to the task addressed in this paper. Also, often indication of importance and dull is embeded into the subtle audio/speech cues and can get lost in transcript, hence to study modality collapse in such scenarios it makes more sense to directly use audio/speech instead of their transcript.

  • this work seems overly narrow in scope without introducing novel findings or methods. the study relies on a single tiny benchmark, MOMENTS, which comprises moments from football matches (146 important and 43 non-important moments). this benchmark is far too small to draw findings with statistical significance or broader applicability.

  • figure 1 taking up an entire page seems redundant. i strongly recommend that the authors familiarize themselves with the standards and expectations of ICLR.

  • l.274 indicates that this work primarily verifies the findings of an earlier paper surikuchi et al. (2026), suggesting a lack of novel contributions. none of the findings appear broad or generalizable, or significant. while the authors highlight a potential "conflict" between modalities (where ablating either modality improves prediction toward the ground truth), they do not propose any solutions to address this issue.

  • the entire paper boils down to a single analysis: computing attention knockout across modalities to measure whether output likelihood increases or drops on a very small evaluation set. an ICLR paper requires a much more substantial technical contribution than this.

Questions for authors

please see above under weaknesses.

1 comment

Many thanks for the review. Here are our responses to the points raised under weaknesses:

  1. In L041-043, we attribute the lack of increase in confidence to modality collapse in the specific setting of Surikuchi et al. (2026), not in general. The moments we use are carefully curated: their importance labels come from highlight reels put together by professional experts, and each moment was selected such that its event type (e.g., a corner) is recognizable from the video and from the commentary separately, and then verified by humans. So, unlike in the example of an object's color and its audio, neither modality lacks relevant cues. Moreover, in this paper we do not rely on confidence: our definition is causal (L180-181), and a modality without relevant cues would have an impact close to zero, which we now distinguish from a modality that misleads the model (see point 2 of our response to Reviewer DozU).
  2. Surikuchi et al. (2026) also evaluated newer models on this task, including Qwen3-Omni-30B-A3B and Qwen3-VL-4B, and found that they do not perform better: among open-source models, Qwen2.5-VL (VL) and Qwen2.5-Omni (Omni) obtain marginally better scores, with no model significantly outperforming the others. So, stronger cross-modal alignment in newer models does not translate to this task, and we focus on VL and Omni because the behavioral claims we examine were made about them.
    • About audio: we agree that speech carries cues (e.g., prosody) that a transcript loses. We use the transcribed commentary because VL does not support audio; this way, both models receive the same commentary input, which keeps the comparison between them controlled. Extending the analysis to the audio input of Omni is a natural next step.
  3. A small correction: the subset has 146 moments in total, 103 important and 43 non-important. We use this subset deliberately, since each moment is manually verified for its event type, which our category-level analysis requires, and each moment is analyzed individually through its own interventions. The main effects are also statistically significant at this size: for example, synergy decays from goals to corners/throw-ins to shots-on-target (VL: 79% → 59% → 39%; Omni: 73% → 33% → 18%; permutation test for trend, p < 0.001 for both). Like we mention under limitations (L447-450), our approach is not specific to this task and can be extended to other datasets to establish how far the findings generalize.
  4. Thank you for this suggestion. We have a more compact version of the figure, which we will use in the revision.
  5. About our contributions (last two points): we respectfully disagree that the paper boils down to a single analysis. Beyond attention knockout, we introduce a per-moment, four-quadrant diagnostic of modality use, validate the knockout against value zeroing (Appendix A) and a width-matched control (Appendix B), and complement it with layer-level analyses of attention share and logit lens (Section 4). This framework is not specific to football and can be applied to other multimodal classification tasks. About mitigation: we deliberately keep verification and mitigation separate. Like we mention in L455-461, mitigation first requires causally localizing where the collapse happens, and the mechanistic techniques available for LLMs cannot be trivially generalized to multimodal models. We are working on this as part of our next steps.

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

Related papers

Discussion (0)

Sign in to comment.