acceptodds
Under review as a conference paper at ICLR 2027

Where Do Spatial Answers Come From? Tracing Utility Across the Visual Scene

Abstract

Vision-language models (VLMs) have made rapid progress on video-based spatial reasoning. However, it remains unclear which visual features these models rely on to generate their answers. Attention maps and representation probes reveal where a model attends and what information it encodes, but do not by themselves establish which feature directions a particular answer depends on. We introduce Utility Probe (U-probe), a training-free method that identifies an answer-conditioned feature subspace, termed Utility Space (U-space), by coupling visual representations with gradients of the answer. Deletion and steering interventions validate the causal role of U-space, whose contributions we trace across the visual scene. The leading feature directions in U-space are interpretable as spatial concepts, and subspaces for corresponding tasks agree across datasets. Deployed on VLMs finetuned on spatial tasks, U-probe reveals that scene-level spatial reasoning is supported by not only the queried objects but also features distributed over the scene, particularly on walls and floors, i.e., scene structure. During finetuning, regional utility after as few as 20 steps predicts later model performance, providing an early diagnostic of spatial SFT, whereas regional attention does not. Our results connect answer-conditioned feature subspaces to their use across the scene and identify scene structure as an important substrate for spatial reasoning.

Primary area:
interpretability and explainable AI
Keywords:
Spatial reasoning, Interpretability, Vision-language models, Feature subspace, Causal intervention

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_MRDiOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors identify an important issue: Vision-language models frequently encounter challenges in explaining which visual cues drive their spatial reasoning capabilities.

And devise an interesting method to address it: Utility Probe (U-probe), a training-free approach that identifies answer-conditioned feature subspaces they call Utility Space (U-Space). By linking token representations with answer gradients, this technique establishes the causal roles of specific visual elements without requiring model retraining.

They then ran comprehensive experiments to demonstrate that scene-level spatial tasks depend heavily on scene structure, such as foundational elements like walls and floors, rather than solely relying on the primary objects mentioned in prompts. Furthermore, they show that tracking regional utility early in the supervised fine-tuning process serves as an effective diagnostic tool to predict final model performance.

Critical strengths

Reading this paper, 2 things stand out for me.

  1. It helps bring to light a classic “AI interpretability issue”: proving causality rather than correlation. While I have seen papers that show where a model looks (a lot of them), this paper focuses on “why it answered the way it did”. That is, I think Utility Probe (U-probe) is very interesting.

  2. It presents a practical "early warning system" for training expensive models: knowing early on whether a training run is going to succeed or fail could be a huge practical advantage. The authors discovered that by measuring how much "utility" a model retains on key background elements (like walls and floors) after just 20 training steps, they can predict how well that model will perform on spatial benchmarks hundreds of steps later.

Critical weaknesses

Reproducibility.

I appreciate that the authors include details about the models and experiments; however, my primary benchmark for reproducibility is an inspectable anonymous codebase. I believe the diagnosis is complex enough that minor implementation choices in gradient hooking, tensor normalization, or token boundary assignment can alter the results. While I may not necessarily run the code, submitting it now will make it easier for the authors to publish it later so others can use it.

Questions for authors
  1. In Figure 3, all the extreme-projection patches look to me like nondescript crops of plain wall texture. Beyond manual inspection, how did the authors quantitatively confirm that these top U-Space axes reflect meaningful 3D geometry rather than visual noise or high-norm background tokens?

  2. U-probe relies on a first-order linear approximation (coupling hidden token states with first-order answer score gradients). How well does this linear framework capture multi-step, non-linear spatial reasoning across 64 video frames? Have the authors tested whether higher-order gradient interactions or non-linear probes reveal different feature subspaces?

  3. Since U-probe pools tokens across all 64 frames and erases temporal sequence, isn't a frame-agnostic linear probe mismatched for diagnosing sequential spatial reasoning?

  4. Any insights on how one could convert U-Space into a performance booster?

3 comments

We appreciate your positive assessment of U-probe's causal analysis and its application as an early training diagnostic. Thank you for highlighting reproducibility and the interpretation of our findings. We now address your questions below.

  • Code release

Thank you for your suggestion. We have submitted an anonymous codebase that could reproduce U-probe, SFT, and predictor analysis on all the models involved in our paper, in Supplementary Material. We will also release the trained model checkpoints to ensure a full reproduction of our work.

  • Quantifying the effect of top U-Space axes

We clarify that we do not assume any pre-defined semantics of the token projection on the U-space axes. We first quantitatively confirm that the U-space axes carry answer-relevant features in Section 3, that they can alter the answer logit and steer the generated answer. Moving the token features along noise (e.g., random directions) does not give the same effect, as shown in Figure 2. We then interpret what are the semantics of the U-space features in the visualizations and logit lens. The extreme-projection patches do not come from a pre-defined semantics that they carry geometry information, but from the fact that they carry much answer-relevant features through U-space.

To investigate the relationship between the U-space projection and the token norm, we compute the Spearman correlation ρ\rho between a token's projection onto the top U-space axis and its norm, over all image tokens. We choose 10 samples from each question type in VSI-Bench. The correlation is weak (∣ρ∣|\rho| median is 0.12 for signed projection value; ∣ρ∣|\rho| median is 0.07 for the magnitude; all results are on 5 VLMs), suggesting that the U-space projection is not choosing high-norm tokens.

To further investigate whether the U-space-projected token features carry geometry information, we decode the depth at that patch position. We use Depth-Anything-V2 to estimate the depth and build a small 3-layer convolutional decoder on the tokens, after projecting them onto the rank-64 U-space. The results are added to Appendix K.4 in the revised paper and show that those tokens carry local scene geometry through U-space.

  • The effectiveness of the linear framework on non-linear spatial reasoning

We agree that U-probe uses a first-order linear approximation, and that spatial reasoning requires a multi-step and non-linear procedure. We therefore evaluate U-space interventions by rerunning the non-linear model after feature perturbations, and Figure 2 shows answer score changes and generated answer effects beyond the matched controls. These results validate the selected features' causal effect rather than relying on the approximation alone. We agree that higher-order gradients or non-linear probe of the answer can reveal complementary features, and we leave investigating them as future work.

  • The pooling effect on sequential spatial reasoning

We agree that pooling across all 64 frames does not retain an explicit frame index. However, it aggregates features and answer gradients computed from the ordered video. These quantities can contain temporal and cross-frame context. Reordering the input frames and recomputing the model can change the features, gradients and resulting U-Space. Thus, the aggregation need not be invariant to input frame order. The "appearance order" interpretations and regional deletion results provide evidence that the method can identify relevant features on a sequential task. We thank you for your suggestion and agree that a frame-aware method might further enhance the capability of U-probe. We plan to investigate it in the future work.

  • Insights on how one could convert U-Space into a performance booster

We believe that U-space could shed light on how to improve VLM performance, and discuss this possibility in pretraining and architectural design. U-space could inform the pretraining to build scene understanding by paying attention to scene structure regions, or motivate new VLM architectures that explicitly perceive the scene structure along with the object-centric representations. We provide a short discussion in Section 5 and Appendices I-J in the submitted paper on this topic, and we plan to investigate them in the future work.

Thanks to the authors for the detailed response and for providing the code. I appreciate the extra effort with the depth decoder in Appendix K.4 and the token-norm checks.

That said, I have to admit I'm a bit disappointed that several concerns are not directly engaged:

  1. Figure 3 remains uninterpretable: If proving that these extreme tokens carry 3D geometry requires training an external 3-layer conv decoder on a rank-64 subspace in the appendix, then Figure 3 itself doesn't demonstrate this to a human reader. I would like to see the intuition and understand the figure.

  2. Temporal collapse and linear assumptions: Re-running the non-linear model to show that perturbing features changes logits only confirms causality. It doesn't show that a first-order linear probe models multi-step reasoning. More importantly, simply noting that the VLM's inputs are sequential doesn't change the fact that U-probe pools all 4,096 tokens across 64 frames into a single static average. It remains frame-agnostic, which sidesteps how dynamic, sequential video reasoning actually unfolds over time.

  3. Actionability and "Future Work": Deferring non-linear probes, frame-aware formulations, and performance-boosting mechanisms to future work confirms that U-Space currently functions as a post-hoc diagnostic rather than an actionable tool for model improvement.

Overall, the paper presents an interesting diagnostic, but the methodological mismatch for sequential video tasks and the heavy reliance on "future work" for key limitations leave me on the fence.

Thank you for your follow-up and for acknowledging the code release and additional analyses. We now address your questions below.

  • The interpretability results

We have revised Figure 3 to show full-frame examples for two question types: counting and room size. We select patches with large positive and negative projections onto the leading U-Space axis. The same three frames are shown for both question types, comparing their visual patterns directly. Positive (P) and negative (N) projections are displayed together within each frame. We choose the same frame across different question types to make the task dependence visible and the visual concept more interpretable in the scene context. "Counting" highlights object-related regions, including object groups, water-tap, and chair parts, in contrast to relatively uniform regions. "Room size" emphasizes extended surfaces and their boundaries or corners. These visual patterns provide qualitative interpretations related to object quantity and spatial extent, illustrating how the features associated with the leading axes depend on the question type. Full patch visualizations across models and question types as well as logit lens interpretation of the language-side U-space axis results are retained in Appendix E. The rank-64 depth decoder in Appendix K.4 provides separate evidence of recoverable depth information and is not required to interpret the visual contrasts in Figure 3.

  • Linear approximation and temporal reasoning.

We appreciate the opportunity to clarify the scope of U-probe. Our goal of using U-probe is to understand what features are used by the VLMs to answer spatial questions, while it does not attempt to reconstruct the complete reasoning algorithm. It does not train a linear probe for the reasoning steps or predefined spatial variables. Its first-order objective identifies intervention directions, whose actual effects are evaluated through the full nonlinear VLM.

We agree that the aggregate matrix does not explicitly reconstruct temporal interactions or individual reasoning steps. The aggregation constructs a shared feature basis from token representations and answer gradients computed on the ordered video. Token-level utility remains associated with each token’s spatial position and frame. Such information is utilized by U-probe in temporal-sensitive spatial reasoning, e.g. the "appearance order" questions, to identify causal features for answering those questions in VLMs.

  • Actionability

We agree that U-probe is not yet an actionable tool for model improvement. We use U-probe to identify and interpret the features used by VLMs in spatial reasoning, to diagnose and analyze the spatial SFT of VLMs, and to build an early warning system for the costly training. As Appendix J explains, an informative ranking across models does not imply that increasing the same utility fraction within one model will improve its accuracy.

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

This paper introduces Utility Probe, a training-free method that combines visual token representations with answer gradients to identify an answer-conditioned feature subspace (U-Space). Deletion, feature injection, and language-side patching experiments test whether this subspace affects spatial answers. The authors then attribute utility to scene regions across five VLMs, finding that wall and floor tokens often carry answer-relevant features alongside queried objects. They also report that changes in regional utility after early spatial fine-tuning correlate with the models’ later vision margins.

Critical strengths
  1. U-probe links token representations to the gradient of a specific answer, then tests the resulting subspace by deletion, injection, and language-side patching. The effects on answer scores and generated answers exceed the matched random-subspace and PCA controls. This makes the proposed subspace more compelling than an attention map or a probe showing only that spatial information is encoded.

  2. The paper traces answer-relevant features across the visual scene. Region-restricted deletion shows that features carried by wall and floor tokens affect several spatial question types beyond the queried-object region, using size-matched random-token controls. The analysis spans five VLMs and two datasets, and early regional utility is more associated with later vision margins than the attention baseline. These results offer a useful, testable account of where answer-relevant visual information is carried, without requiring the stronger assumption that each token represents only its local image content.

Critical weaknesses
  1. I am unsure how independently the interventions validate U-Space and its regional interpretation. My understanding is that, on the data used to construct U-Space, its leading positive directions maximize the first-order decrease in answer score under deletion. Were the subspaces in Figure 2a fitted on questions separate from the 50 evaluation questions? If not, how much does the comparison with random directions and feature PCA test beyond that first-order construction? For Figure 4, could a contiguous-region control help distinguish the wall and floor masks from scattered random tokens? Could those tokens carry global information rather than the local scene geometry attributed to them? These distinctions matter to how strongly I can interpret the intervention results.

  2. I would like to understand how broadly the early diagnostic has been validated. With only five independent models in the permutation test, how sensitive is the step-20 result to the choice of models or task-specific region rules? Does the weaker step-20 SPAR result support the same predictive claim? Could Gemma-4 be compared with the other models at the same training step and under comparable token-handling conditions? Finally, has the difference between utility and attention been tested directly, rather than inferred from their separate significance tests? Answers to these questions would help me assess whether this is a useful diagnostic for the studied models or a more general predictor.

Questions for authors

Clarifications

  1. In Figure 2a, is U-Space fitted on the same 50 questions used to measure the deletion effect? Is it constructed using the gradient of the same answer whose score is evaluated (FgtF_{\mathrm{gt}} or FownF_{\mathrm{own}})?

  2. When was the rule assigning each question type to a scene-structure or object region fixed? How were the changes to that rule for SPAR and the multi-view tasks decided?

Additional evidence that would affect my assessment

  1. In Figure 4, does deleting U-Space features from wall and floor tokens remain more damaging than deleting features from size-matched contiguous regions elsewhere in the image? This would help me distinguish a scene-structure effect from an effect of deleting a large connected region.

  2. How stable is the step-20 predictor across different samples of the 100 training questions and different choices of the five models? Has its predictive performance been compared directly with the attention baseline using a paired statistic? These results would help me judge the strength and generality of the early-diagnostic claim.

4 comments

We appreciate your careful assessment of the intervention design and the robustness of the early diagnostic. We address your questions below.

  • U-Space fitting, evaluation, and the first-order objective

We clarify that the questions used for identifying U-space and testing the effect of U-space are different in Section 3. For Figure 2 (a)(b)(c), we choose 50 questions for each type on VSI-Bench to identify the U-space and test on 50 held-out questions, and there is no scene overlap. The question type remains the same, i.e., we use the U-space fit on "room size" questions for testing on the "room size" questions. We use FgtF_{gt} here because we want to identify the feature corresponding to the ground truth answer. Note that either using FownF_{own} or using the same questions during U-space fitting and testing does not change the conclusion, and the effects are stronger in certain cases. For experiments in Section 4, we use FownF_{own}. In Section 4.1, the questions used for fitting the U-space and testing the regional importance are the same.

The leading U-space directions maximize the estimated first-order deletion effect, while this does not guarantee a large effect on a change in the generated answer. Higher-order terms can change the response, and a score change need not translate to the model’s answer change. The experiments in Section 3 therefore measure actual changes in answer scores and generated answers, providing empirical evidence of U-Space’s causal influence under interventions.

  • The assignment rule used in Section 4.2

We highlight that the rule assigning each question type to a semantic region is decided before observing any predictor results and fixed since then. The SPAR rule is designed after we observe that the top-projection patches on U-space almost always contain the dot marker region, regardless of the questions. We therefore hypothesize that marker is important for the VLM to correspond the questions to the scene, as the marked out object could re-appear in the scene and its identity has ambiguity in the question. For example, the question could mention "chair" while there are many different instances of chairs in the scene, and the marker is important for identifying the question-relevant object. More details are given in Appendix H.3 in the submitted paper. The multi-view SPAR experiments are added to further validate the rule, and the rule is also fixed before any results are obtained: if the question is about the room or about where things stand in it, it is assigned scene structure; if it is about the property of the named objects, it is assigned queried object. This rule is stated in Section 4.2 in the submitted paper, and Section 4.1 contains quantitative evidence supporting it in that scene structure matters on those scene-demanding questions. The SPAR rule also applies, that is, when there is a dot/point marker on the object in question, assign queried object of that specific instance and all of its re-appearance across the scene. With the rule fixed for both SPAR video and multi-view datasets, the predictor shows a positive trend that becomes clearer later in training, as shown in Appendix H.3. The SPAR video results in Appendix H.3 thus follow a similar trend as in Figure 5, i.e. the VSI predictor results, supporting the same predictive claim.

  • Continuous region deletion control

We appreciate this suggestion. We first clarify that the scene structure region is not necessarily a large connected region. A typical frame contains the scene structure region separated by furniture, as shown in Figure 4(a). We now quantify the connectivity of the regions. We define a piece of scene structure in a frame as a max group of structure tokens in which any two tokens are connected by a path of steps between tokens adjacent up, down, left or right. Diagonal step does not count. The size ss of a piece is the number of tokens in it. Pieces are found separately in each frame and never join across frames. We find that scene structure regions have 1.96 pieces (std 0.37) on average over 50 scenes. A comparable number occurs on the "others" region, which has 1.99 pieces (std 0.50). The two regions have similar total size, and we then investigate whether they have similar piece size expectations over tokens, i.e, how large a piece a deleted token typically lies in. This is calculated by: (1) Identify pieces 1,...J1,...J from the visual input; (2) Calculate size of them s1,...sJs_1,...s_J; (3) Calculate the expected size of the piece containing a token drawn uniformly from the set, denoted as sˉ\bar{s}. Such a token lies in piece jj with probability sj/∑i=1Jsis_j/\sum_{i=1}^{J}s_i, and thus sˉ=∑j=1Jsj2/∑j=1Jsj\bar{s}=\sum_{j=1}^{J} s_j^2 / \sum_{j=1}^{J} s_j. The sˉ\bar{s} has a median of 23.5 of scene structure and 24.7 of the others region. We therefore use the "others" regions for the control for region connectivity, as it shares similar connectivity statistics with the scene structure region. We do not prevent the token selection from selecting scene structure tokens in order to satisfy the equal-size requirement, but we constrain the selection to not adding the pieces to ensure the connectivity. We perform the same experiments as in Section 4.1, yet we do not observe significant changes of the results. The cell numbers change at most 0.05 with many cells stay the same. This is because many of tokens chosen in our previous control already lie in the "others" region, so the effect is already partially measured. The region deletion results remain similar under the connectivity control we tested. We have added this experiment to the Appendix K.2 in the revised paper.

  • The robustness of the predictor and including Gemma4 in it

We clarify that our predictor results report standard deviation as error bars in Figure 5(b), over 3 non-overlapping 100-question datasets. We have added the clarification to the main paper. To test the robustness of the predictor to different choices of the five models, we perform model leave-one-out (LOO) experiment, where we drop a model and calculate the 4-model predictor result. We have added the results to Appendix K.1. The aggregate correlation remains positive when any model is omitted across all steps, suggesting that it is robust to the choice of the models.

We further compare Gemma-4 with other VLMs at the same early checkpoints. Gemma4 are nearly always ranked the last by our regional utility predictor at every step, and its step-800 vision margin follows it. Adding the model to our correlation calculation changes the step 10 to 50 results to: 0.453(0.007),0.673(0.006),0.624(0.003),0.763(0.003),0.771(0.001)0.453 (0.007), 0.673 (0.006), 0.624 (0.003), 0.763 (0.003), 0.771 (0.001), suggesting that including Gemma4 actually strengthens our conclusion. We do not perform feature resize on Gemma4 as we do on the other 5 VLMs because the training is unstable. Instead, we use the soft-token-method that comes with it to ensure the training stability.

  • Paired statistics between utility and attention baseline

We appreciate your suggestion to compare the two predictors directly with a paired statistic. For each SFT step, question type and subset, we compute the Spearman correlation as in Section 4.2, separately for utility and attention, and take their difference. We average this difference over the three 100-question subsets and apply a two-sided paired tt-test across the 7 question types. The mean differences (pp-values) from step 10 to step 50 are 0.560.56 (0.0480.048), 0.490.49 (0.0120.012), 0.300.30 (0.0100.010), 0.260.26 (0.0350.035), 0.270.27 (0.0280.028). The mean difference is also positive on each of the three subsets at all five steps (0.140.14 to 0.670.67). We can conclude that the utility-based predictor is quantitatively better than the attention-based one. We have added this result to the paper in Appendix K.3.

  • Local scene geometry carried by tokens

We examine whether the scene structure token carry local geometry information by decoding the depth at that position. We use Depth-Anything-V2 to estimate the depth and build a small 3-layer convolutional decoder on the tokens, after projecting them onto the rank-64 U-space. The results are added to Appendix K.4 in the revised paper and show that those tokens carry local scene geometry through U-space.

Thank you for the response and the new experiments. The held-out fitting in Section 3 resolves my concern about circularity, and the connectivity control and the direct utility–attention comparison are useful additions.

Some concerns remain. The depth-decoding result shows that scene-structure tokens encode local geometry, but not that the answer uses it. The paired t-test uses question types as the unit of analysis, whereas Appendix H.2 treats the five models as the experimental units. The early diagnostic is weaker on SPAR, and Gemma-4 is an extreme case.

Given the resolved concerns, I have raised my score by one point. As this paper is outside my area of expertise, my confidence remains 1.

Thank you for acknowledging that the revisions and experiments resolve your earlier concerns, and we appreciate your reconsideration of the score. We address your questions below.

  • Depth information and interpretability of U-space

We agree that depth decoding does not establish that the answer uses the decoded geometric information by itself. Our causal evidence comes from Section 3 and Section 4.1. Appendix K.4 separately establishes that depth information is recoverable from U-Space projections. The interpretation of the semantics of U-space axes is in Section 3.4, and we have revised Figure 3 to add qualitative interpretations through selected patches with high and low projections onto U-space axes, shown in their scene context to demonstrate the two contrasting visual patterns of the visual concept carried by those U-space axes. For example, the "Room Size" axis highlights contrasts between broad surfaces and corners/boundaries, suggesting an association with the concept "area". The full patch visualizations and logit lens further provide interpretations of the U-space axes across different models and question types.

  • Paired statistics

We agree that our previous paired t-test treats the question type as the unit instead of the model, which shows that utility has a higher predictive correlation than attention on average across question types. We instead use an exact paired permutation test using the five models as experimental units. Appendix H.2 shuffles YY across models to test the correlation between early utility and final vision margin. For the paired comparison here, we keep each model paired with its own outcomes and enumerate the possible swaps of its utility and attention figures. Under the null that utility and attention predictor numbers are exchangeable within each model, we enumerate all 252^5=32 swaps, keeping all question types and subsets of each model together and fixing the YY. For each configuration, we recompute the mean difference in Spearman correlations between utility and attention and compute the number that the correlation is larger than the observed one, in absolute value. The floor of this two-sided pp is thus 2/32=0.06252/32=0.0625. We recompute the p-values for those observed differences and get: 0.560.56 (0.1250.125), 0.490.49 (0.06250.0625), 0.300.30 (0.06250.0625), 0.260.26 (0.06250.0625), 0.270.27 (0.06250.0625). Thus, utility ranks the five models more accurately on average at each checkpoint. We have added the results to Appendix K.3, clarifying the difference between the two tests.

  • Generalization to SPAR and Gemma

We agree that the SPAR predictor is weaker than the one on VSI-Bench, and we provide a short discussion in Appendix H.3. The Gemma result is an extreme case in that the model does not behave well, which is diagnosed by our predictor. The primary five-model analysis in Section 4.2 excludes Gemma.

Reviewer_7uvFOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The authors analyze spatial reasoning in VLMs, discovering a subspace in the residual stream that is causally responsible for the VLM’s answer, which they call the U-space. The method for finding the U-space for a set of tasks is training-free, label-free, and counterfactual-free: for layer l, they first-order approximate the rest of the model after layer l using gradients, and then they theoretically solve for the directions which, when removed from the residual stream vector, lead to the largest change in the output logits; these directions, which span the U-space, turn out to be the top eigenvectors of a symmetrized average of outer products of the residual stream vectors and the gradients. An interpretation of this matrix is that the residual stream vector part measures if the direction is captured in the content, and the gradients measure if the direction is causally read by the output. After discovering the U-space, the authors verify it using casual interventions, checking that ablation and steering change the answer. In addition, they qualitatively check on the top eigenvector direction using the logit lens and the top-similarity token patches, finding that it encodes meaning related to the question type that the U-space was derived on. Next, the authors further localize the most important tokens for the U-space to be the background tokens (floor and wall), even when the visual task is not related to the background. They also show that the U-space. can be used to predict performance of a model SFT’d on the visual tasks. For their quantitative evaluations, the authors use 5 VLMs and 2 datasets of question and answer on videos.

Critical strengths
  1. Although the math behind extracting the U-space is not new, the authors are the first to apply it to spatial reasoning in VLMs and to use it for causal interventions. To the best of my knowledge, the technique is not a standard part of the mechanistic interpretability toolkit for discovering feature directions and subspaces in the residual stream, and it has advantages over more common techniques like DAS, linear probes, and diff-in-means by not requiring training, paired counterfactuals, or labels. It is surprising that the authors do not make this comparison themselves in the paper.

  2. The paper demonstrates that their methods hold generally among VLMs and visual QA tasks, showing that the U-space is meaningful in all 5 VLMs tested, and showing that the U-spaces found for corresponding question types are similar across the two datasets, so that they’re not picking up on dataset quirks.

Critical weaknesses
  1. The paper overclaims their novelty in terms of the contribution of “Spatial reasoning runs on scene structure” (line 88). The authors cite Cui et al. 2026 as revealing spatial information encoded in representations, but say that “These observations alone leave the answer’s dependence on that information unresolved” (line 46). Cui et al. 2026 uses causal interventions to prove that their discovered features in the background of the images causally affect the answer, addressing both the causal point as well as the scene structure point.

  2. The SFT results are about predicting the relative rankings of VLMs after SFT, rather than improving SFT performance or predicting an exact accuracy value. It’s not clear why this prediction is useful or how it offers greater understanding of what the VLM is learning during SFT.

Questions for authors
  1. [further experiments] For the figure 2d experiment on the vision component of the language tokens, I think it’s not guaranteed that the subspace being patched is purely in the vision modality, and hence the result of the answer changing may be because the text modality information about the changed prompt is being patched over. Could you instead, for the pairs being patched, hold the prompt constant and reflect the image to change the intended answer instead? This way, if the patching over the language tokens still changes the answer, we know that it had to come from vision information encoded into language tokens. There should be some control measuring the base rate decrease in performance due to analyzing mirrored images.

  2. [further experiments] The section on SFT seems to imply that SFT increases the fraction of the total utility which belongs in the localized wall and floor tokens. It would be interesting to see a plot of how this utility fraction varies during SFT training (i.e. the x-axis is training step).

  3. [further experiments] How did the authors choose the layer to use U-space on? Is there some kind of sweep over layers? Since U-space uses a linear approximation, I would imagine that U-space performs worse for earlier layers.

I already gave the author an accept rank but might adjust it if the authors do not address the suggested experiments in the questions and fix the overclaiming on visual information being stored in the scene background,

6 comments

We appreciate your recognition of U-probe's methodological advantages and your suggestions for clarifying its novelty, practical value, and experimental validation. We address your concerns and questions below.

  • Novelty relative to Cui et al.

We agree that Cui et al. establish a causal role for ordering information carried in background tokens, and we now discuss this result separately from analyses of encoded spatial representations. We have revised the Introduction, Related Work, and contribution statement accordingly, framing our scene-structure contribution as quantifying task-dependent regional utility. These changes are marked in red.

  • Comparison with other interpretability tools

We appreciate the suggestion to make this comparison explicit. Supervised linear probes fit property decoders; DAS optimizes representation alignments against interchange targets; difference-in-means requires contrasting example groups. We have now clarified how the requirements of U-probe differ from those of existing tools in the Introduction and Related Work, marked in red.

  • Why the prediction is useful

The predictor supports early model selection for spatial SFT, allowing further training to be prioritized before expensive training finishes. In Section 4.3, the step-20 utility ranking already identifies Qwen3.5 as the preferred candidate for longer training, consistent with the best-performing model after 800 steps. Therefore, the same model selection decision can be made much earlier. On 4 Nvidia GH200, it means saving more than 200 hours of SFT on the 5 models. Appendix H.5 further illustrates its application as an early warning of weak VSI-Bench performance in the tested Gemma4 model and settings. Appendix H.4 reports weaker correlations with later vision margins for training loss and training accuracy, supporting utility as a more informative selection signal in our experiments. We have added this clarification at the end of Section 4.2 in the revised paper. We provide a short discussion in Section 5 and Appendices I-J in the submitted paper on improving accuracy using these insights, and we plan to investigate them in the future work.

  • Further experiments 1: Language-side U-space

We appreciate your idea on making this experiment more sound. We design a new experiment following your suggestion, that uses the same prompt with reflected images. We pick those QA pairs that the model can answer correctly both before or after the reflection, where the reflection would reverse the answer accordingly. For example, it can contain a frame where the two queried objects coexist and the answer will reverse if the frame is mirrored. We use GPT-6-Astra to build the reflection pairs and validate the answer manually. Processing mirrored image barely decrease the model performance in the relational direction questions. The change of model accuracy on reflected images on 100 randomly sampled left/right questions is within [-5%,2%]. We use GPT-6-Astra to judge if the reflection changes the ground truth answer. Note that this test is just for demonstrating that the performance change is small, instead of performing the language-side U-space patching experiment. We perform the same experiment as in Section 3. The rank-64 language-side patch retains 0.92 of the full-patch flip effect, compared with 0.02 for its complement and 0.00 for a matched-rank random subspace, consistent with the original experiment. We have added the new results in Appendix C.2 in the revised paper, marked in red. Because the prompt is held fixed, the transferred difference is caused by the visual input rather than a change in question wording. We acknowledge that the language-side U-space remains to be an approximation of the vision information in the language tokens due to the complexity of token features, as we claim in the text that we "isolate the vision content approximately". We nevertheless show the effect of the stored vision information in our empirical validations.

  • Further experiments 2: Utility fraction during SFT

We provide the visualizations of utility fraction during SFT in Appendix H.6 in the revised paper, with x-axis being the training step and y-axis being the utility fraction in percentage. Different models have different trends in the utility fraction change, and our predictor utilizes that change at each step.

We clarify that we do not claim the utility fraction will increase during SFT. Instead, the scene structure utility fraction can decrease during SFT, as reported in Section 4.2. The predictor uses relative changes across models: on scene-demanding tasks, models that retain more of their utility fraction on scene structure tend to have larger final vision margins. Because this is a normalized allocation, a decreasing fraction can imply a relative increase elsewhere. It does not imply that scene-structure information has disappeared.

  • Further experiments 3: The layer choice for U-probe

We agree that this is an important detail to add, and it can also contribute to the understanding of the working mechanisms of VLM. The layer choice originates from an observation of a tension inside VLMs, that the deletion cost of image-side U-space and the patching effect of the language-side U-space has a trade off, and at 0.6 depth (the working layer in our experiments) these two effects are both preserved well. We thus use the layer closest to 0.6 of model depth as a common heuristic. Appendix A.2 in the revised paper now provides sweeps of image token deletion, concentration of the leading directions, language-side patching, and sensitivity to blocking text-to-image attention. Earlier layers can suffer from poor linear approximation and thus have low deletion effects, while later layers can become less sensitive to changes in image tokens or to further visual reads. Around the chosen depth, both image-side deletion and language-side patching retain measurable effects. Together, the 0.6 depth is chosen for every model and used as a heuristic for all involved VLMs.

Thank you to the authors for the thorough rebuttal. The rebuttal satisfies my earlier concerns about novelty relative to Cui et al, the language-side U-space experiment, and the layer choice for U-probe. Please remember to reference Figure 24 from the text of Appendix H.6.

I’m still not convinced of the usefulness of the scene structure utility fraction during SFT experiments. Since the claim is not that the fraction of scene structure utility goes up during SFT, it seems that the evidence for usefulness hinges on predicting the relative ordering of only 5 models’ performance after SFT, which is a very small sample; the models may also have significantly different base performance gaps (Qwen3.5, 3, and 2.5), making this prediction task easier. I therefore maintain my weak accept score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.