acceptodds
Withdrawn from ICLR 2027

Causal Analysis of Physical Understanding in Visual Representations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Decoding physical variables from visual representations does not establish how their encoded information contributes to task prediction. The challenge is to separate target-specific contributions from general intervention effects and the loss of information shared with related variables. We introduce CausalProbe, a framework for representation-level causal analysis. It progressively erases target-predictive directions while protecting selected readouts of related variables. Random and reverse-role erasure provide comparisons for interpreting the resulting task effects. Across six frozen encoders, erasing second-order geometry impairs grasping-affordance prediction on SuctionNet much more than erasing first-order geometry, supporting its stronger causal contribution. On Physion++, spatial erasure similarly affects future-contact prediction more than motion erasure. Removing either family without protection also reduces the other’s predictability, indicating overlap between spatial and motion representations. Our framework moves beyond identifying what visual representations encode to examining how that information supports task prediction.

Primary area:
interpretability and explainable AI
Keywords:
Representation Probing, Subspace Intervention, Functional Attribution, Geometric Representation, Physical Reasoning

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_GL5SOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper introduces CausalProbe, a framework for representation-level causal analysis via subspace erasure for vision models

Critical strengths

The paper tackles a very interesting and timely problem of understanding the 'world models' inside vision models. Particularly, moving from correlative to causal analysis is super important and the paper addresses this head on.

The two tasks (suction affordance and future contact prediction) are sufficiently diverse and also targeted in my mind. And the second-order dimensionality result is impressive and interesting! A 10 dimensional subspace is sufficient for removing grasping affordance.

I also found the experimental procedure detailed yet sound. The control vs erasure iterative procedure is a solid methodology in my opinion.

In general, I am quite positive about the paper but have a few concerns that, if addressed, would raise my score.

Critical weaknesses

I am a bit concerned about the mixed results for separating out spatial and motion information. I appreciate the authors being upfront about the entanglement of the two, which is to be expected, but it is important to have a more detailed analysis and leverage additional existing metrics along this line to better understand how the information is tangled; across layers, neurons, etc. For example, Kowal et al (2022) provide a metric for doing so that could help quantify this is a more finegrained manner. It would be ideal to relate task degradation to achieved target suppression in the results here to get a sense of the specific trade-off.

It remains unclear how much Physion++ contact prediction depends on temporal information, since late-frame spatial cues may already support the task (e.g., via camera motion blur or spurious correlations in the data). A useful experiment would compare the full-video input with the last visible frame repeated across the input sequence, retaining the same target–zone cue and retraining the contact head under identical splits. Similar AUROC would suggest that temporal history contributes little in this setup, limiting what the weak motion-erasure effect reveals about the models’ physical reasoning.

[a] Kowal et al. A Deeper Dive Into What Deep Spatiotemporal Networks Encode: Quantifying Static vs. Dynamic Information (CVPR 2022)

Questions for authors

The following are two extensions; either of which would significantly strengthen the paper:

Could we discover the physics subspaces? Existing work has discovered multidimensional subspaces in vision models. (https://github.com/goodfire-ai/block-sparse-featurizer) BSFs applied to the model, rank subspaces based on their alignment with probes, then remove. This would show we can discover these physical variables unsupervised to some degree.

Diff of means vs. probes. The authors only use probes, which are good for detection (filtering). But since decoding weights can differ from the activation patterns associated with a variable, it would strengthen the attribution results to repeat representative interventions using an alternative direction estimator, such as Haufe-transformed patterns or an appropriate mean-difference/forward-encoding baseline [b].

[a] Fel et al. Structuring Sparsity: Block-Sparse Featurizers Capture Visual Concept Manifolds (Arxiv 2026) [b] Haufe et al. On the interpretation of weight vectors of linear models in multivariate neuroimaging. (NeuroImage 2014)

1 comment
AuthorsOpenReview ↗

Response to Reviewer GL5S (part 2)

2. Quantifying entanglement and relating suppression to task loss — Critical Weakness 1

“it is important to have a more detailed analysis and leverage additional existing metrics along this line to better understand how the information is tangled; across layers, neurons, etc.”

We added quantitative subspace-overlap measurements and an intermediate-layer analysis. Section 5.3.3 and Fig. 9 compare normalized principal-angle overlap between the two families’ decoder subspaces, using rank- and dimension-matched random subspaces as references. Position–velocity overlap exceeds first-/second-order geometry overlap in all six encoders, including after chance adjustment (Appendix J). This provides quantitative evidence consistent with the stronger constraints that protection imposes in the position–velocity experiments.

We additionally analyze V-JEPA 2 blocks 8, 16, and 24 for both the original spatial/motion descriptors and the revised position/velocity variables (Appendix K, Fig. 56). Position prediction improves across these blocks, whereas velocity prediction peaks at the intermediate block. Position–velocity decoder overlap is nonmonotonic and remains above the random-orientation range throughout. These results distinguish changes in predictability from changes in subspace alignment. The measurements concern predictive subspaces rather than individual-neuron selectivity.

“It would be ideal to relate task degradation to achieved target suppression in the results here to get a sense of the specific trade-off.”

Figure 8 now directly plots task loss against achieved target-prediction loss. Each point corresponds to an erasure rank, with protected and unprotected trajectories overlaid for each variable and encoder. This separates the amount of information suppressed, as measured by the probes, from the number of erased directions.

For position erasure, greater suppression generally accompanies greater task loss. For velocity erasure, protection restricts target suppression in every encoder and task loss in five; VGGT exhibits the opposite task-loss ordering. Together with the control-prediction curves in Fig. 5, these comparisons clarify when greater task impairment accompanies both stronger target suppression and collateral effects on the related variable.

Reviewer_DJY1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces CausalProbe, a framework for examining how physical-variable information in frozen visual representations supports downstream prediction. It iteratively removes target-predictive directions while preserving selected control-variable decoders, and evaluates targeted, reverse-role, and random interventions, together with restoration experiments. Experiments cover six encoders on SuctionNet seal-quality prediction and Physion++ future-contact prediction. The principal findings are that second-order geometry erasure substantially affects seal prediction, while spatial erasure generally affects contact prediction more than motion erasure under the implemented protocols.

Critical strengths
  1. A well-motivated analysis with useful controls. The paper addresses an important distinction between decoding physical variables and assessing their contribution to downstream prediction, providing substantially more informative evidence than baseline probing alone.

  2. The SuctionNet evidence is comparatively strong, providing a coherent empirical finding within the evaluated setting.

Critical weaknesses
  1. The Physion++ comparisons do not sufficiently establish relative contributions of spatial and motion information. Both protected interventions achieve incomplete suppression. Maximum spatial test R2R^2 remains 0.821-0.886 at the common ranks in Figure 35, while maximum motion R2R^2 remains 0.758-0.814 after reverse erasure (Sections C.8-C.9). Target dimensionalities, protected spans, and terminal ranks also differ. Consequently, the larger spatial-erasure effect establishes an asymmetry between the implemented interventions, but does not yet establish a comparable ordering of the variable families' contributions. This limits the strength of the video study's principal interpretation.

  2. The incremental methodological contribution is insufficiently isolated. The one-shot comparison in Figure 26 targets only three matrix coordinates and removes three directions, whereas the iterative method targets the full eight-group family and removes 9-13 directions. The comparison therefore changes target coverage and intervention size alongside iteration. Moreover, prior work such as SPLINCE already studies concept removal with preservation of related label information. Its covariance-preservation constraint differs from the fixed-decoder constraint here, but this relationship deserves substantive discussion to establish the new analytical capability provided by CausalProbe. Holstege et al., NeurIPS 2025

Questions for authors
  1. Clarification affecting the empirical interpretation: How much of the spatial-motion asymmetry remains at overlapping removed ranks, accompanied by component-level suppression measurements? Which evidence supports a relative-contribution interpretation beyond differences in the achieved interventions?

  2. Clarification affecting the methodological contribution: What distinct capability does fixed-decoder protection provide relative to existing preservation-aware erasure approaches? Can comparisons using the same complete target family and comparable preservation constraints isolate this capability from target coverage and removed rank?

3 comments
AuthorsOpenReview ↗

Response to Reviewer DJY1 (part 1)

Thank you for recognizing the value of distinguishing variable decodability from its contribution to task prediction, and for highlighting the SuctionNet evidence. We have revisited temporal-history dependence, revised the main video experiment, and expanded the comparisons between protected and unprotected interventions. Below, we address your weaknesses and questions separately.

Weakness 1: Interpretation of the video results

“The Physion++ comparisons do not sufficiently establish relative contributions of spatial and motion information.”

To investigate the weak motion-erasure effect, we tested temporal-history dependence in Physion++’s object-contact prediction task. Replacing the RGB history with repeated last frames, while retaining the original query cue and separately refitting the prediction heads, yields comparable contact-prediction performance across the six encoders (Appendix H). This indicates that the evaluated models can reach their original performance without the full RGB history, limiting what this task reveals about motion contributions.

We therefore revised the main video experiment to future-position prediction using the same Physion++ videos and splits. Across six encoders and the main 30-, 40-, and 50-frame horizons, repeated frames reduce average task (R2R^2) from 0.619 to 0.455, demonstrating a clear benefit from temporal history (Appendix I).

The revised interventions show that erasing either position or velocity impairs future-position prediction beyond random erasure while protecting the other variable’s designated readouts. Without protection, erasure also affects prediction of the other variable, particularly when velocity erasure disrupts position prediction. Our revised conclusion is that both variables contribute to future-position prediction and have coupled representations, rather than that spatial information universally contributes more than motion.

AuthorsOpenReview ↗

Response to Reviewer DJY1 (part 2)

Weakness 2: Family-based targeting and methodological contribution

Purpose of the single-descriptor comparison

“The comparison therefore changes target coverage and intervention size alongside iteration.”

This experiment evaluates family-based targeting, rather than isolating iteration alone. Constructing a family of physically grounded descriptors is a central design choice in our framework. The change in target coverage is intentional: we ask whether targeting one descriptor sufficiently suppresses the task-relevant information associated with the underlying physical variable.

Different descriptors of the same geometry can be accessible through different predictive directions. For example, a linear probe may predict a nonlinear descriptor of the second-order matrix, such as its Frobenius norm, even when the matrix entries are poorly predicted. Erasing one descriptor’s predictive subspace therefore need not suppress other task-relevant expressions of the same geometry.

Appendix B.11 compares independent single-target erasures, each using repeated probe fitting and the same first-order-family protection, with family-level erasure. Appendix B.12 adds the one-shot matrix reference. These experiments motivate targeting and auditing the descriptor family jointly, rather than equating suppression of one descriptor with suppression of the underlying physical variable. Their purpose is not to establish that iteration alone is superior at an identical erasure budget.

Relationship to preservation-aware concept erasure

“Moreover, prior work such as SPLICE already studies concept removal with preservation of related label information.”

Our primary contribution is an analysis framework for investigating how physical-variable information contributes to task prediction, rather than a new general-purpose concept-erasure algorithm. The protection mechanism is designed for this analysis: progressively erase target-predictive directions while holding selected related-variable readouts unchanged. Its constraint and construction differ from the cited method, and our experiments demonstrate its intended behavior within the framework.

The distinctions are:

  • Motivation. SPLINCE—the name used in the published paper—studies concept removal while preserving related label information, including applications to model interpretation. Our framework examines how information about a physical variable contributes to a distinct prediction task under explicit related-variable protection. The task response itself is not protected.

  • Method. SPLINCE uses a minimum-distortion oblique projection with feature–label covariance-preservation constraints. Our method uses iterative orthogonal erasure constrained to preserve selected decoder predictions for every example. We then evaluate target prediction, control prediction, and task performance along the intervention trajectory. Holstege et al., NeurIPS 2025

  • Findings. Geometry permits substantial target suppression with related-variable prediction largely maintained. Position and velocity exhibit stronger coupling: protection restricts target suppression, whereas unprotected erasure introduces collateral prediction losses. The framework therefore identifies both task contributions that persist under protection and situations where overlap limits selective attribution.

Dear Reviewer, we would appreciate any feedback on whether our response has addressed your concerns. We are happy to clarify any remaining questions. Thank you for your time.

Reviewer_BYyvOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

In this paper the authors propose CausalProbe, which extends previous representation analysis work allowing for task dependence on different encoder features to be elicited. The main contribution of the paper is extending a feature erasure method to selectively remove target features from encoders without impacting a control set of features. This allows them to disambiguate the effects of different feature spaces that would have previously been confounded in previous feature erasure methods. The authors apply their method to a suction affordance task and a future contact prediction task, demonstrating different levels of task impact using their method on different feature sets.

Critical strengths

This paper takes a very natural next step extension to a previous method for understanding the impact of encoder features on task performance. They are able to demonstrate that their method clearly works in that they can protect variables while subtracting out others in multiple application examples. The paper in general is well written!

Critical weaknesses

The impact of this work is not super clear. The authors demonstrate that their method can determine sets of features which uniquely change the task performance with some brief comments that these features they find to be more impactful than others are intuitive. However, its not clear if there were any insights  gained from doing this analysis?

This method could use a clearer direction on what its purpose is on these tasks- is it to ensure representations are encoding known proper robust features?

The way that “causal” is used in this paper is probably better substituted for “unique”. The authors are pulling out only the unique impact of a feature rather than the full causal impact of a feature - which is what previous methods this paper is based on already do.

I also am unclear whether the methodology as proposed is conceptually an improvement over a method that does not use a protection subspace. Imagine we want to measure the impact of two variables on task performance. If two variables are correlated, then they naturally will share a subset of encoding features that can predict both and may also have a unique set of features which can predict the non-shared components of the variables. Because of this, removing a single one of these variables from the encoder features naturally will impact the decodability of the other. It's not clear that this is unwanted behavior, this measures the full impact of removing a feature. With the authors’ method, any direction that decodes the other variable would be protected, so shared features would not be subtracted out. This means that the full effect of a variable on task performance is not accounted for in their method. Its unclear why this method is better than taking one variable, subtracting out its encoding features and measuring task impact, doing the same for the other feature and then comparing the task performance drop. One thing that would improve this work would be to show the shared feature impact (the space where both tasks can be decoded) as well as the unique impact currently explored - that way the full task impact of a variable could be decomposed into unique and shared components.

Minor point: The notation for x being the input and X being the features is confusing. These should probably be different variable names.

Questions for authors

Are there any results here that would not have been obtained by looking at the impact of features individually without the protected subspace? I would guess that you could obtain the same results (2nd order features impact more than 1st order etc) by looking at the relative task impact of each feature set individually. In any case, this would be a needed baseline to see if this proposed method is able to do things that the previous method without protection subspaces could not.

2 comments

Thank you for your constructive review and for recognizing the clarity and soundness of our experimental procedure. Your feedback prompted us to clarify the scientific questions and substantially expand the empirical analysis. In the revised manuscript, protected and unprotected erasure are complementary parts of CausalProbe, rather than treating protection as a universally superior replacement for unprotected erasure. We address your comments individually below.

Critical Weakness 1: Scientific insights and motivation

The impact of this work is not super clear. The authors demonstrate that their method can determine sets of features which uniquely change the task performance with some brief comments that these features they find to be more impactful than others are intuitive. However, its not clear if there were any insights gained from doing this analysis?

Our central question is not simply which physical variables can be decoded, or which variables are physically expected to matter. It is which information learned by visual encoders actually supports downstream prediction, and how that information is organized across related physical variables. The revised experiments reveal two contrasting patterns:

  • SuctionNet (Figs. 3, 5, and 9). Across all six encoders, second-order geometry erasure causes substantially larger grasping-affordance degradation than first-order erasure. This holds with and without protection, and the strong second-order effect persists while first-order readouts are preserved. Restoration of second-order information further improves task performance. This establishes a strong, selective functional dependence in the learned representations, beyond the physical intuition that curvature matters for suction.
  • Physion++ (Figs. 4, 6–8). We now use future 3D position prediction as the main video task (retaining future-contact prediction in the appendix). Current position and velocity both support future prediction, but their predictive decoder subspaces are substantially more aligned than the geometry families. Their erasure produces asymmetric cross-variable effects, and joint unprotected erasure drives mean future-position R2R^2 below zero across all six encoders. These results reveal a more strongly coupled organization of task-relevant information.

Together, the two settings show that physical information differs not only in task relevance, but also in representational coupling—a distinction that probing accuracy alone cannot establish.

This method could use a clearer direction on what its purpose is on these tasks- is it to ensure representations are encoding known proper robust features?

The purpose is functional analysis of existing representations, not verification that models encode a predefined set of “proper” features. Probing establishes the availability of information; interventions test how removing that information affects task prediction. Comparing protected and unprotected interventions further reveals whether the effect persists when related-variable readouts are preserved. We have made this research objective explicit in the revised Introduction and contributions.

Critical Weakness 2: “Causal” versus “unique” contribution

The way that “causal” is used in this paper is probably better substituted for “unique”. The authors are pulling out only the unique impact of a feature rather than the full causal impact of a feature - which is what previous methods this paper is based on already do.

We agree that protected erasure primarily captures the unique contribution of target-variable information under the constraint of preserving related-variable information, rather than its full causal contribution. In the revised manuscript, we emphasize the jointly analysis of protected and unprotected erasure. While unprotected erasure measures the overall task impact of removing target-predictive information, protected erasure isolates the effect that persists when related-variable readouts are preserved. Combining these complementary interventions provides a more comprehensive understanding of how physical information contributes to downstream tasks, including the influence of shared representations across related variables. We have updated both the methodology and experimental analyses accordingly.

Dear Reviewer, we would appreciate any feedback on whether our response has addressed your concerns. We are happy to clarify any remaining questions. Thank you for your time.

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

Related papers

Discussion (0)

Sign in to comment.