acceptodds
Under review as a conference paper at ICLR 2027

Knowing Without Deciding: The See–Decision Gap in Vision-Language Models

Abstract

Vision-language models (VLMs) can internally see the correct spatial relation yet still give the wrong answer. We term this the Encoding–Decision Gap: the ground-truth relation remains reliably decodable across many consecutive layers, but fails to determine the model's final response. Combining spatial readouts with causal interventions, we trace the gap to a dissociation between spatial encoding and decision control. Spatial information is broadly recoverable across intermediate states, yet causal influence over the answer is concentrated in a small subset of token–layer states, forming an encoding-to-decision bottleneck. Moreover, the directions that most strongly control the answer extend substantially beyond the relation-defined spatial subspace, while remaining relation-specific. We therefore use the model's own intermediate spatial estimate to select which relation to strengthen in the intermediate spatial states, and let the remaining computation propagate this change to the output. Without supplying an external answer or new spatial information, this intervention reconnects information already present in the model to a causally effective pathway, improving spatial reasoning across multiple datasets and model families. Our results show that spatial reasoning failures can stem not from missing representations, but from a failure to transfer encoded information into the decision, making encoding-to-decision transfer a concrete target for mechanistic analysis and intervention in VLMs.

Primary area:
interpretability and explainable AI
Keywords:
interpretability, mechanistic interpretability, vision-language model, multimodal, spatial reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_LvMBOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This work aims to uncover the mechanisms driving spatial reasoning in VLMs, and address why a model might derive accurate spatial relationships in its latent space but fail to accurately answer questions about it.

Critical strengths
  • Identifying the underlying mechanisms behind spatial reasoning in Vision Language Models is a timely and important problem.

  • The related works exploration is thorough, and background on the line of research is well covered.

Critical weaknesses
  • I have some novelty concerns with this manuscript. The introduction states that the paper contains 3 novel contributions. Contribution 1, the identification of the encoding-decision gap, was already stated and explored in at least two papers cited by this submission: "Linear Mechanisms for Spatiotemporal Reasoning in VLMs" [1] and "Knowing Isn’t Always Saying: When Do Spatial Encodings Reach Answers in Vision-Language Models?" [2]. ([2] is recent, but if the work was concurrent, it should clearly be acknowledged as such.) Contribution 2 mainly just restates Contribution 1, with a small mechanistic explanation addition. Therefore, half of the novel contributions of this work (as stated in the introduction) appear to not be novel, and the authors seem to be aware of the preexisting works that tackle this.

  • It also appears that Contribution 2 and 3 contradict each other. Contribution 2 states that much of the causal mechanism behind spatial reasoning lies outside of the "relation-defined spatial subspace". However, the recovery method in S3.4 relies on these latent space directions for steering. So it seems like the submission is arguing that the latent space steering is a poor method for control, but all results use that exact approach to steer output.

  • The paper claims that a few text latents form the causal bottleneck behind the encoding-decision gap, but doesn't show that this is a particular problem for wrong answers, while not occurring for correct answers. Without such evidence, it appears that this is just an emergent characteristic of such models, and does not serve as the explanation behind the gap.

Minor Weakness:

  • Fig 4 and 5b has overlapping text. Also, Fig 6 appears corrupted somehow?

[1] https://arxiv.org/abs/2601.12626 [2] https://arxiv.org/pdf/2608.22916

Questions for authors
  • The proposed method for increasing spatial QA accuracy is upper bounded by readout accuracy, and seems to loop until readout-level performance is reached. Why wouldn't one just use the readout as the answer?

  • Were different data splits chosen to analyze adequate steering layers vs. performing the steering, or were they the same?

2 comments

Thanks for pointing the clarity issue in the introduction, particularly in distinguishing our contributions from recent related work. We thank the reviewer for the constructive feedback and address each concern below.

W1: Novelty and distinction from prior work.

We agree that previous studies have made important progress in understanding spatial representations and their causal roles in VLMs. We also acknowledge that the gap between correct intermediate spatial representations and incorrect final predictions is not a new observation. Our work builds on these findings and takes a closer look at the internal mechanisms that connect spatial encoding to final decisions.

Linear Mechanisms for Spatiotemporal Reasoning in Vision Language Models investigates how spatial information is represented and used in VLMs. It identifies spatial IDs associated with object-word representations and explains their role in spatial predictions through a linear mechanism.

Knowing Isn't Always Saying focuses on when and where spatial information reaches the final answer. It examines how this process depends on layers, token positions, and prompt formats, and uses causal interventions to study the conditions under which spatial representations affect model outputs.

Our work asks a different set of questions. We want to understand what happens inside the model between spatial encoding and the final decision, and which internal states are responsible for carrying the causal influence of spatial information. Although previous studies have shown that spatial representations can influence model outputs, this does not fully explain how spatial information contributes to decision-related signals or how its influence passes through downstream states.

We investigate these questions through two mechanistic analyses. First, we study the spatial-to-decision transition by comparing intermediate image-conditioned spatial representations with late decision-related signals. This allows us to examine how the role of spatial information changes across different stages of the model, beyond simply observing the gap between intermediate readouts and final predictions. Second, we investigate the internal causal pathway through which spatial information affects the final decision. Through targeted interventions and Block/Keep experiments, we identify the downstream text states that mediate the effects of upstream spatial interventions.

These two analyses address different parts of the mechanism. The first examines the relationship between spatial representations and decision-related signals. The second identifies the downstream states through which spatial information exerts its causal influence. We also use steering to test whether the identified spatial representations can influence the model's own generation, which provides functional validation of our mechanistic findings.

We will revise the introduction and contribution statements to explain these differences more clearly.

W2: Spatial subspace and decision-related signals.

We thank the reviewer for raising this concern. We believe there may be a misunderstanding of our findings. We do not claim that latent space steering is ineffective or that spatial representations have little causal influence on model outputs. Our steering experiments actually demonstrate that interventions on spatial representations can change the model's final generation.

Our subspace analysis examines the relationship between spatial representations and decision-related signals. We find that the decision-related gradients are not strongly aligned with the spatial subspace. This distinction is part of the encoding–decision gap we investigate: the spatial information encoded within the model is not necessarily reflected in the same way in its decision-related signals.

This does not contradict our steering results. The gradient overlap measures the local sensitivity of a specific decision-related signal, whereas steering tests the effect of spatial interventions on the model's actual generation. The two experiments examine different aspects of the model's internal mechanisms, and a low overlap does not mean that spatial interventions cannot affect the final output.

W3: Interpretation of the causal bottleneck.

We thank the reviewer for raising this concern. We apologize if our original description caused confusion. We do not claim that the causal bottleneck is specific to incorrect predictions or that the encoding–decision gap is the cause of prediction errors.

Our encoding–decision gap describes a mechanistic phenomenon: spatial representations and decision-related signals are not fully aligned, and spatial information undergoes changes as it contributes to the final decision. Our goal is to understand this process and the internal mechanisms involved, not to attribute incorrect predictions to the gap itself.

Our causal bottleneck analysis addresses another part of this process. Through causal interventions and Block/Keep experiments, we identify a small number of downstream text states that play important roles in transmitting the influence of spatial information to the final decision. This pathway can exist in both correct and incorrect predictions.

Prediction errors may have different causes, including problems at the initial stage. Therefore, the existence of an encoding–decision gap or a causal bottleneck does not by itself explain why a particular prediction is incorrect. We will clarify these distinctions and add the relevant analysis of initial-stage errors to the revised manuscript.

Q1: Why not use the intermediate readout directly?

We thank the reviewer for this question. Our goal is to understand how spatial representations influence the model's final generation, not to develop a method for improving spatial QA accuracy.

A direct readout extracts an answer from intermediate representations, but skips the subsequent decision process that we want to study. Our steering experiments examine whether interventions on these representations can change the model's own generation.

The readout tells us what spatial information is available in the model, and steering helps us test how this information affects its final output. We will make this distinction clearer in the revised manuscript.

Q2: Selection of intervention layers.

We thank the reviewer for this question. We selected the intervention layers based on our layer-wise mechanistic analysis, not by searching for the layers that achieved the highest accuracy.

We conducted this analysis across multiple datasets and found that the same model consistently showed the same important layers, regardless of the dataset used. In other words, even if we select the intervention layers using a different dataset, we arrive at the same choice. This suggests that the selected layers reflect a consistent property of the model's internal mechanisms and are not specific to the evaluation dataset.

We will clarify the layer-selection procedure and report this cross-dataset consistency in the revised manuscript.

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

The paper studies spatial-relation errors in vision-language models for which an intermediate readout predicts the correct relation but native generation does not. It constructs relation directions from real-minus-gray object-pair representations on Synthetic-3D, analyzes the transition from spatial to answer-option organization, and identifies influential text states using a relation-conditioned gradient score. Ablation, signed interventions, and mediation experiments examine how spatial changes affect the output. The recovery procedure uses an intermediate reader to choose a target relation and iteratively modifies object-pair states until native generation agrees with that target. Evaluation covers six VLMs and three datasets, with improvements reported for all 18 model–dataset combinations. The proposed contribution is a mechanistic account of the encoding–decision gap and a method for exploiting intermediate spatial information.

Critical strengths

The paper goes beyond correlational probing. Main-text Tables 2–3 show signed effects of spatial interventions and selective mediation through ranked states. For example, for L23→L25, blocking the Top-7 states retains 22.1% of the signed effect versus 95.6% for random states, while keeping the Top-7 retains 91.2% versus 1.5%. Supplementary Section C.3 also tests positive and negative real-minus-gray interventions. These experiments provide useful evidence that the identified states participate in transmitting spatial information.

The readout–generation discrepancy has useful cross-domain and behavioral evidence. Relation directions are estimated from a separate synthetic source, and Supplementary Section B.2 tests horizontal-flip equivariance rather than relying only on unperturbed probe accuracy. Table 4 reports recovery gains across several model families and all three target datasets. These results make intermediate spatial representations a worthwhile target for further analysis.

Critical weaknesses

The experiments establish causal influence more clearly than they establish the proposed cause of the failures. In Section 3.3, states are ranked using a ground-truth-conditioned objective, and the principal Drop/Keep experiment zeros entire hidden states. This can identify states needed for answering without establishing that a defective encoding-to-decision transfer specifically causes the final-wrong cases. The signed interventions in Supplementary C.3 act on the same Top-7 states selected with the ground-truth-conditioned score, so their wrong-to-correct outcomes are not independent of label information. Moreover, the reported analyses do not contrast this mechanism between matched final-correct and final-wrong examples with correct readouts. The directional argument has a related limitation: the four centered class means span at most three dimensions, so the fact that most gradient energy lies outside their span is not sufficient to establish a pathological mismatch. Table 1's random-token control compares token locations, not the overlap expected for equally low-dimensional subspaces. The evidence supports a spatially relevant pathway, but the stronger claim that its bottleneck/misalignment explains the errors needs a discriminating control.

Recovery success does not yet distinguish pathway repair from enforcing the reader's prediction. Equation 6 explicitly optimizes target-versus-competitor margins, and Supplementary A.5 repeats the update and checks generation until it matches the reader, allowing up to 64 steps. Table 4 does not compare the same optimization with norm- and compute-matched alternative directions/sites or report intervention size and cost. In addition, Supplementary A.3 selects the Qwen2.5-VL-3B reader using 80 COCO-two examples. The final head ranking uses a label-free causal-RMS criterion, but the 23-head candidate pool is described only as heads with "a strong image-dependent spatial signal," while the same section refers to "previously measured COCO-two spatial readout accuracy" for these heads; the construction of that pool, whether label-dependent measurements influenced it, overlap with the reported evaluation, and corresponding reader-selection details for the other models are not sufficiently specified.

Questions for authors

Among examples with correct readouts, does the transfer effect differ between final-wrong and matched final-correct cases? How does the spatial-subspace overlap compare with a dimension-matched control?

How does recovery change with norm- and step-matched alternative directions or token sites under the same reader targets? Please also report correct-to-wrong transitions, perturbation norms relative to native states, and typical step counts.

How was the candidate-head pool built, and which reader head is used for each model? Several Table 4 values are not attainable as integer counts over the stated sample sizes (e.g., 83.01% over 440), and Supplementary B.1 reports 141 errors on a COCO-two subset, more than any Table 4 baseline makes on all 440 examples; which model, samples, and prompt format were used?

2 comments

We thank the reviewer for their careful review and helpful feedback. We address the questions and concerns below.

W1: Transfer effects in final-correct and final-wrong examples.

We conducted an additional experiment comparing final-correct and final-wrong examples with correct intermediate spatial readouts. We matched examples with similar intermediate spatial representations and examined how spatial interventions affected their final decisions.

We found that both groups responded similarly to spatial interventions. We also examined how image-induced signals affected the final decision margin and found both positive and negative effects in both groups. However, examples with negative effects tended to have lower initial decision margins, making them more likely to produce incorrect final predictions.

These results suggest that final prediction errors are not necessarily caused by weaker spatial-to-decision transfer. Even when the model correctly represents the spatial relation and responds to spatial interventions, the final prediction can still be incorrect. The initial decision margin also matters: when the margin is small, a negative contribution from visual information is more likely to change the final answer. We will include these findings in the revised paper.

W2: Spatial subspace and decision-related signals.

We agree with the reviewer that low spatial-subspace overlap alone is not sufficient to support our original interpretation. The spatial subspace has a much lower dimension than the full hidden space, so a low overlap does not necessarily indicate weak information transfer. We will revise this interpretation in the paper.

We conducted an additional experiment using dimension-matched random subspaces. The results show that the spatial subspace has approximately two to three times higher overlap with decision-related gradients than the random baseline. This indicates that the spatial subspace has a non-random relationship with decision-related signals, despite its low absolute overlap.

To further examine whether the recovery improvement depends on the identified spatial directions, we conducted a random-subspace recovery experiment using the same experimental settings as in the paper. We replaced the spatial directions with random directions of the same dimensionality, while keeping the reader targets, intervention positions, and perturbation constraints unchanged. We also matched the number of optimization steps individually for each example. Random-subspace recovery achieved only 71.59 accuracy, slightly below the 72.05 baseline, showing no meaningful improvement over the baseline. In contrast, recovery using the identified spatial directions substantially improved accuracy. This provides further evidence that the recovery improvement depends on the identified spatial directions rather than arbitrary low-dimensional interventions.

We also examined whether recovery disrupts initially correct predictions. Spatial recovery resulted in 34 correct-to-wrong transitions. The average number of search evaluations was 2.77 across all 440 examples, or 7.62 among the 160 examples requiring intervention. The average residual perturbation L2 norm was 1.86. These statistics characterize the intervention cost and the changes introduced by recovery.

We will revise our interpretation of the overlap analysis and include these additional controls and statistics in the revised paper.

Q1: Transfer effects and spatial-subspace controls.

Thank you for these questions. We have addressed the transfer effects in W1 and the spatial-subspace controls in W2, where we present the additional experiments and discuss their results.

Q2: Recovery controls.

Thank you for raising this concern. Recovery uses the intermediate reader prediction as its target, but does not directly replace the model's final output. To examine whether the improvement simply comes from forcing the model to follow the reader, we conducted an additional random-subspace recovery experiment under the same reader targets and matched optimization steps. We present the results and additional recovery statistics in W2.

Q3: Head selection and numerical inconsistencies.

Thank you for pointing out these issues. We would like to clarify the head selection procedure and explain the two numerical inconsistencies.

For head selection, We apologize for this typographical error. The size of the candidate set should be 20 rather than 23, and we will correct this in the revised manuscript. We use a two-stage procedure. First, we evaluate all candidate heads on a subset of 80 examples from the dataset and select the top 20 based on their causal effects on downstream spatial representations. We then evaluate these 20 heads on the full dataset and select the single best head with the highest downstream spatial RMS. The spatial directions are learned from Synthetic dataset. Ground-truth labels from the target dataset are not used to select or rank the heads.

For the first numerical inconsistency, the reported 83.01 readout accuracy for Qwen2-VL-2B on COCO-two was mistakenly copied from an earlier Controlled-B experiment. We re-evaluated the model on all 440 COCO-two examples and obtained a corrected readout accuracy of 70.91%.

The second inconsistency concerns the 141 final-wrong examples reported in Supplementary B.1. In our earlier experiments, we extracted the first spatial relation mentioned in the generated response. However, this could lead to incorrect answer parsing. We later changed the parser to extract the last spatial relation but overlooked this earlier experiment when preparing the supplementary material. We apologize for this oversight.

We are re-running the analysis on Qwen2.5-VL-3B with the corrected parser. The updated results show [N] final-wrong examples. Among them, [M] have a positive average intermediate spatial margin, which corresponds to [X]%.

We will correct Table 4, Figure 2, and the supplementary material. We will also clarify the head selection and evaluation procedures in the revised paper.

Reviewer_zF3nOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper reveals that "the spatial information is broadly recoverable across intermediate states, but the causal influence over the final answer is concentrated in a small subset of token-layer states". The authors then leverage the model's own intermediate spatial estimate to select which relation to strengthen in the intermediate spatial states. Intervening on spatial information in intermediate layers improves the model's spatial reasoning.

Critical strengths

I am not an expert in this field. I vaguely remember seeing existing papers on revealing the gap between the representation level of the model and the model's final outcomes (possibly some ACL or previous NeurIPS or ICLR paper), but I do believe that the authors have made another step in terms of revealing that the causal influences of such divergence remain localized to the intermediate layers and show that we can steer the model's behavior, at least for spatial reasoning (more of the QA form), on the synthesized dataset.

Critical weaknesses

What is the novelty of the work compared to the general behavior steering line of research? Is this work a combination of mechanistic analysis and steering? How do you position your work among these two lines of research?

The authors test their idea primarily on synthetic datasets with a controlled setup. What about real-world noisy datasets such as GQA? Does the intervention still work?

I generally have this concern about the word "reasoning"; at least for the specific controlled setup, it is in the form of QA to test the model's understanding of spatial relations. Does the model really reason? Could the authors test some recent visual reasoning models and see how the intervention affects the reasoning traces of these models?

Questions for authors

Please see the weakness points.

1 comment
AuthorsOpenReview ↗

Response to Reviewer

Thanks for pointing the clarity issue in the introduction, particularly in distinguishing our contributions from recent related work. We thank the reviewer for the constructive feedback and address each concern below.

W1: Novelty and distinction from prior work.

We agree with the reviewer that prior studies have extensively investigated spatial representations in VLMs and their influence on predictions. For example, "Linear Mechanisms for Spatiotemporal Reasoning in Vision Language Models" identifies spatial representations and examines their causal effects, while "Knowing Isn't Always Saying" investigates when and where spatial information influences model outputs. "Seeing is not Enough" further highlights the gap between visual perception and decision-making.

However, "how spatial representations are transformed into decision-related signals and how their causal influence propagates through downstream internal states remain insufficiently understood."

Our work addresses these gaps through two complementary contributions. First, we characterize the "spatial-to-decision transfer", examining how intermediate image-conditioned spatial representations differ from late decision-stage signals and what internal changes occur during this transition. Second, we identify the "downstream causal mediation pathway", revealing the text states through which upstream spatial representations influence final decisions. Unlike prior work that primarily examines spatial encoding or when and where spatial information influences predictions, we investigate how this information is incorporated into and propagated through the decision process.

Finally, activation steering provides functional validation of our mechanistic findings by demonstrating that the identified spatial representations can influence model generation. "The primary novelty of our work is uncovering how spatial information transitions into decision-related signals and identifying the downstream internal pathways mediating its influence on model decisions."

W2: Generalization to more complex visual reasoning tasks.

We agree that evaluating a broader range of visual reasoning tasks could further demonstrate the generality of our findings. However, our study focuses specifically on spatial relations, aiming to understand how relational information is represented and transferred into decision-related signals within VLMs.

Our experiments include both synthetic images and real-world COCO images. COCO contains complex natural scenes with multiple objects and diverse spatial arrangements, providing a realistic setting for investigating spatial-to-decision transfer. We believe these experiments provide sufficient evidence for the spatial relation tasks studied in this paper. Although GQA covers a broader range of question types, this does not necessarily make it a more suitable benchmark for our specific mechanistic analysis.

More broadly, our approach is not necessarily limited to spatial relations. Since we analyze different relations through relation-specific directions in the model's hidden representations, a natural extension would be to investigate whether similar representation and transfer mechanisms exist for other types of object relationships. We see this as a promising direction for future work and will clarify it in the revised manuscript.

We agree that spatial relation questions do not necessarily require explicit multi-step reasoning. Our goal is not to evaluate general reasoning capabilities, but to understand how spatial relations are internally represented and utilized by VLMs to inform final decisions.

Specifically, we investigate spatial relation cues encoded within the LLM component and how manipulating these representations can improve final predictions without fine-tuning the LLM. Spatial relation QA provides a controlled setting to examine the connection between spatial representations and model decisions.

We appreciate the suggestion to explore reasoning traces in recent visual reasoning models. While this could extend our analysis to explicit reasoning processes, our current focus is on internal representations and their causal effects. Activation steering provides functional validation by demonstrating that intervening in spatial representations can improve final generation.

We will clarify this scope and distinguish spatial relation recognition from broader spatial reasoning in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.