acceptodds
Under review as a conference paper at ICLR 2027

Editing What the Video Contributes: Visual Intervention Meets Fisher Geometry

16 views

Abstract

Successful editing in Video-LLM can mask Visual Blindness: the correction may rely on text-side shortcuts rather than videos. We study this ambiguity through a structure-preserving Visual-Null intervention, revealing that substantial editing gains persist without informative video. We translate this behavioral contrast into Fisher geometry to isolate editing directions beyond the blind response. The resulting Fisher-orthogonal residual defines video-grounded editability, capturing both separation from the Visual-Null response and editing target controllability. This formulation yields VIGOR (Visual-Intervention Guided Orthogonal Residualization), a closed-form, minimum-information update that improves the video-conditioned objective while preserving the Visual-Null objective to first order. We prove that greater video-grounded editability lowers the minimum functional cost of a prescribed target improvement, informing principled layer selection within the locate-then-edit paradigm. Across four Video-LLMs on VMEB, VIGOR achieves the highest average editing score among evaluated methods, combining near-perfect reliability with near-zero Visual-Null gains. It also largely preserves general video understanding and representation structure after 100 sequential edits. Our code is provided in the supplementary materials; all data will be released soon on Github.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Video Large Language Models, Model Editing, Visual Grounding, Fisher Geometry

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_u6ycOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VIGOR, a Fisher-geometric model editing framework designed to mitigate Visual Blindness in Video-LLMs. The method introduces a structure-preserving Visual-Null intervention to identify editing gains that persist without informative video content. It then applies Fisher-orthogonal residualization to remove the blind editing component and derives a closed-form, minimum-information update. The authors also introduce Grounded Interaction Editability for principled layer selection. Experiments across four Video-LLMs on VMEB demonstrate strong editing performance, near-zero Visual-Null gains, and improved stability under sequential edits.

Critical strengths

1. Relevant problem and clear motivation.\textbf{1. Relevant problem and clear motivation.} The paper addresses Visual Blindness in Video-LLM editing, where successful editing may rely on text-side shortcuts rather than visual evidence. The Visual-Null intervention provides a useful way to distinguish editing success from video dependence.

2. Principled geometric formulation.\textbf{2. Principled geometric formulation.} The proposed Fisher-orthogonal residualization connects Visual-Null analysis to parameter updates. The minimum-information objective and Grounded Interaction Editability provide a coherent framework for editing direction and layer selection.

3. Comprehensive experimental evaluation.\textbf{3. Comprehensive experimental evaluation.} Experiments across four Video-LLMs demonstrate strong editing performance and near-zero Visual-Null gains. The ablation studies and sequential editing experiments further support the effectiveness of the proposed components.

Critical weaknesses

1. Insufficient justification for choosing Fisher Geometry.\textbf{1. Insufficient justification for choosing Fisher Geometry.} The paper motivates Fisher Geometry by arguing that Euclidean distance does not adequately reflect functional changes in model predictions. While the connection between Fisher Information and local KL divergence is well established, the paper does not sufficiently explain why Fisher Geometry is particularly suitable compared with other non-Euclidean metrics. The ablation study only compares Fisher Geometry with Euclidean geometry, leaving the benefits over alternative curvature-based metrics unclear.

2. Limited methodological novelty over prior work.\textbf{2. Limited methodological novelty over prior work.} Visual Blindness has already been investigated in Video-LLM editing, while orthogonal projection and Fisher-based layer selection have also been explored in existing model editing methods. Although combining these ideas for video-grounded editing is meaningful, the paper does not clearly establish which aspects of the proposed formulation represent fundamentally new contributions beyond existing approaches.

3. Limited validation of the theoretical functional cost.\textbf{3. Limited validation of the theoretical functional cost.} The relationship between Grounded Interaction Editability and minimum functional cost is derived under a local approximation. However, the paper does not directly establish how well the theoretical Fisher cost corresponds to actual changes in model predictions under finite or sequential updates. Additional analysis comparing the predicted cost with measured distributional changes would strengthen the theoretical claims.

4. Visual-Null suppression does not necessarily establish semantic video grounding.\textbf{4. Visual-Null suppression does not necessarily establish semantic video grounding.} Near-zero Visual-Null editing gains indicate reduced dependence on the blind pathway, but do not necessarily demonstrate that the edited model correctly utilizes the semantic or temporal information in videos. More controlled counterfactual evaluations would strengthen this claim.

Questions for authors

1. Fisher Geometry:\textbf{1. Fisher Geometry:} Why is Fisher Geometry particularly appropriate for video-grounded editing compared with other non-Euclidean or curvature-based metrics? Could the authors provide further theoretical justification or comparisons with alternative metrics?

2. Methodological novelty:\textbf{2. Methodological novelty:} Given existing work on Visual Blindness, orthogonal projection, and Fisher-based model editing, what is the most distinctive technical contribution of VIGOR? Could the authors clarify its differences from closely related approaches?

3. Theoretical cost validation:\textbf{3. Theoretical cost validation:} How closely does the theoretical minimum functional cost correlate with actual KL divergence between the original and edited models? Does this relationship remain consistent under sequential edits?

4. Video grounding:\textbf{4. Video grounding:} Does suppressing Visual-Null gains necessarily imply that the edited model relies on meaningful visual information? Could the authors evaluate semantically controlled counterfactual videos, such as changes in actions or temporal order, to further validate video-grounded editing?

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

This paper investigates Visual Blindness in Video-LLM knowledge editing, where successful edits may rely on text-side shortcuts rather than video evidence. To address this issue, the authors introduce a structure-preserving Visual-Null intervention to distinguish video-conditioned editing gains from those achievable without informative visual content. Building on this observation, they propose VIGOR, a Fisher-geometric editing framework that removes the Visual-Null response through orthogonal residualization and derives a closed-form, minimum-information parameter update. The authors further define Grounded Interaction Editability to characterize the video-specific editing capacity across layers and guide editing layer selection. Experiments on the VMEB benchmark across four Video-LLMs demonstrate that VIGOR achieves strong editing performance while substantially suppressing Visual-Null gains. Additional ablation studies and sequential editing experiments suggest that the proposed method effectively balances editing reliability, locality, and preservation of general video understanding capabilities.

Critical strengths
  1. Meaningful problem formulation. The paper addresses an important limitation in Video-LLM knowledge editing: successful answer correction does not necessarily indicate reliance on visual evidence. The proposed Visual-Null intervention provides a useful way to identify editing gains that persist without informative video content, highlighting a limitation of conventional editing evaluation.

  2. Principled methodology with strong empirical results. VIGOR combines Fisher-geometric orthogonal residualization with a minimum-information editing objective, providing a mathematically interpretable framework for suppressing blind editing gains while limiting functional disruption. Experiments across four Video-LLMs demonstrate strong editing performance and substantially reduced Visual-Null gains. The ablation studies and sequential editing experiments further support the effectiveness of the proposed components and their ability to preserve general model capabilities.

Critical weaknesses

Practical motivation and broader generalization could be further discussed. While the paper demonstrates strong editing performance and robustness to video perturbations, further discussion of real-world applications would help clarify its practical significance. Additionally, evaluating whether editing effects transfer to semantically similar but distinct videos would provide further insights into the generalization capabilities of the proposed method.

Questions for authors
  1. Further discussion of practical applications and potential generalization to semantically similar but distinct videos would help clarify the significance of Video-LLM editing.

  2. Since the proposed Fisher-orthogonal residualization constrains the editing direction, it would be useful to investigate whether this constraint increases the number of iterations required for convergence, along with its impact on editing efficiency.

Reviewer_T5k2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies visual blindness in Video-LLM knowledge editing, where an edit may succeed by strengthening text-side associations instead of relying on video evidence. The authors introduce a structure-preserving Visual-Null intervention to identify improvements that persist when informative video content is removed. They then formulate video-grounded editability using Fisher geometry and propose VIGOR, which projects away the blind editing response and computes a minimum-information update. Experiments on four Video-LLMs show strong editing performance and markedly reduced Visual-Null gains.

Critical strengths
  1. The distinction between successful editing and genuinely video-dependent editing is meaningful and underexplored.

  2. Fisher-orthogonal residualization provides a coherent framework for separating video-conditioned editing directions from the Visual-Null response.

  3. The study includes four Video-LLMs, multiple baselines, and ablations of Visual-Null residualization and Fisher geometry.

Critical weaknesses
  1. The model could become sensitive to visual-token presence or input-distribution differences without using the relevant video semantics. Although alternative null constructions are examined, a stronger evaluation would pair the same question with semantically different videos requiring different answers.

  2. The analysis preserves the Visual-Null objective to first order, whereas practical editing uses finite updates and multiple optimization steps. The accuracy of this approximation for the final edited model requires more direct validation.

  3. Teacher-forced target-token accuracy may overstate success in natural generation. Free-form response evaluation under contradictory video evidence would strengthen the conclusions.

  4. The inverse relationship between the defined editability and the theoretical minimum functional cost follows from the formulation itself. Actual predictive KL divergence, interference, and long-term preservation would offer more independent evidence.

Questions for authors

NA

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.