acceptodds
Under review as a conference paper at ICLR 2027

Analyzing Spatial Cognition via Behavioral Interpretability

Abstract

What are LLMs learning to solve spatial cognition tasks? Interpretability techniques usually inspect a model's verbalized reasoning or internal activations, but chain-of-thought traces can be unfaithful and activations are inaccessible for closed models. Behavioral psychology faces the same two obstacles, since self-reports are unreliable and the brain is hard to observe, and instead studies the mind by modeling people's choices and mistakes. We propose to study LLMs the same way by synthesizing symbolic programs that predict their choices and mistakes. On four interactive spatial cognition tasks, our evolutionary search finds programs that match each model's action distribution to within 0.07 nats while being two orders of magnitude simpler than a memorization baseline. The programs reveal that the strongest models perform shortest-path planning in maze navigation, raster scanning and spatial memory-based pruning to find treasures in the Cambridge spatial working memory task, and lookahead planning over visual waypoints for egocentric navigation. On the other hand, weaker models perform greedy goal search in mazes and get stuck in dead ends, exhibit poor retention in spatial working memory and reopen boxes known to be empty, and oscillate in place for egocentric navigation.

Primary area:
interpretability and explainable AI
Keywords:
Spatial Cognition, Frontier Models, Behavioral Interpretability

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_qQ7POpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The authors carefully designed a series of experiments to evaluate the spatial cognition capabilities of large language models (LLMs). Drawing on behavioural psychology and computational simulation, the study examined planning, working memory, and egocentric navigation mechanisms. The findings suggest that leading models substantially outperform mid-tier and smaller models. Their behaviour extends beyond random or naive moves and is consistent with the use of internalised, algorithm-like heuristics.

Critical strengths

The authors designed the experiments very carefully. In particular, the framework works on closed models with no need of weight access, internal activations. By fitting against execution traces, the synthesised code explains why models fail rather than assuming an optimal solver. Uses situational forward Kullback-Leibler divergence to normalise for model entropy and cross-validates continuous/discrete parameter bounds via coordinate descent.

Critical weaknesses

In 3D egocentric navigation, programs do not process raw images directly; instead, camera views an approximated top-down 2D occupancy grid, creating a representation gap between VLM visual inputs and program inputs.

Synthesized programs describe what behavioral rules reproduce the model's choices, but do not explain how these heuristics are physically represented within internal transformer attention heads or circuits.

Data and code are not available.

Questions for authors

Why use LLMs whose source codes are not avaible, such as GPT 5.4, Anonymous-6rhc4, Gemma 4 31B, Qwen 3.6 27B, Qwen 3.5 4B, and InternVL 3.5 38B? Using opensource LLMs will be better to guarantee the conditions of the experiments.

Are data and code available? How can we reproduce the results?

2 comments
AuthorsOpenReview ↗

Official author response to Reviewer qQ7P (1/3): Representation gap in 3D egocentric navigation

1. Representation gap in 3D egocentric navigation

In 3D egocentric navigation, programs do not process raw images directly; instead, camera views an approximated top-down 2D occupancy grid, creating a representation gap between VLM visual inputs and program inputs.

Egocentric navigation is particularly important to analyze now because recent frontier VLMs exhibit a sharp rise in spatial competence (Section 1). These tasks test multi-turn decision-making, perceptual grounding, and visual navigation (Section 2)—central emerging capabilities whose behavioral strategies are worth characterizing. At the same time, we agree that the representation gap between raw egocentric RGB images and top-down 2D grids is a genuine challenge.

To narrow this gap, we use the 3D scene and camera poses to reconstruct visibility-masked local occupancy maps with explicit occlusion, landmark metadata, and fitted perceptual limits such as field-of-view, depth, and waypoint subsampling constraints (Appendices B.1 and B.2). Because this reduces rather than eliminates perceptual mismatch, we scope our claims to geometry- and action-level behavior rather than raw vision. Even under this approximation, evolved programs fit egocentric behavior far better than the best seed hypothesis: on Route Retracing and Shortcut Discovery they raise next-action accuracy from 41.9% to 56.2% and from 42.2% to 55.8%, and lower situational KL from 0.31 to 0.05 and 0.07 nats (Section 4.2 and Appendix A.3). The gains hold on unseen 3D scenes, with test accuracy of 49.7% vs. 38.2% and 59.2% vs. 44.5% (Appendix A.2).

Area_Chair_YwbCOpenReview ↗

Feedback

Hi, the review deadline is approaching. Please take a look at the authors' rebuttal and post your feedback (and update your rating if appropriate) at your earliest convenience. Thank you! AC

Reviewer_p9PXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes using synthesized symbolic programs from LLM behavior as a method to intepret the policies learned by LLMs. They use evolutionary search to find close symbolic programs, and identify patterns and capability differences in models across four spatial cognition tasks.

Critical strengths
  1. The core concept of using evolutionary search over symbolic programs to interpret LLM policies is novel and useful.
  2. The tasks chosen are interpretable and the comparisons across input modalities is useful.
Critical weaknesses

I am a bit skeptical of the Situational KL Divergence metric, since the grouped situations may hide useful differences. Essentially, low Situational KL Divergence could just mean that the grouping is too coarse and not informative enough. Can the authors please expand on this?

Questions for authors

Please look at the Weaknesses section above.

Reviewer_s4qBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how behavioral interpretability can be used to analyze the spatial cognition strategies of large language models (LLMs) and vision-language models (VLMs). The authors propose an LLM-driven evolutionary program synthesis framework that iteratively generates, modifies, and optimizes symbolic programs to model target models' action choices and errors. Experimental results show that the synthesized programs can closely approximate model behavior distributions while maintaining low complexity. The authors further use these programs to analyze differences in path planning, memory utilization, and navigation strategies across models.

Critical strengths
  1. Meaningful research problem. The paper analyzes model decision-making strategies through observed behavior, offering a valuable approach to understanding black-box models without access to their internal parameters. Compared with task success rates alone, behavioral modeling provides additional insights into decision preferences and typical errors.

  2. Methodological novelty and interpretability. The authors combine LLM-driven evolutionary search over program structures with parameter optimization to construct executable symbolic models, providing a flexible framework for analyzing complex sequential decision-making behavior.

  3. Broad experimental coverage and valuable behavioral insights. The experiments cover multiple models, spatial cognition tasks, and input modalities. The results demonstrate a favorable trade-off between behavioral fidelity and program complexity. The authors also identify differences in planning strategies between text and visual inputs and analyze behaviors such as memory-based pruning, local greedy decisions, and repeated actions.

Critical weaknesses
  1. Although the paper evaluates several predefined behavioral strategies, its comparisons are mainly limited to seed programs, random baselines, and a memorization baseline. Direct comparisons with other interpretable behavioral modeling methods, such as decision trees, rule-learning methods, or alternative program-induction approaches, are missing. This makes it difficult to fully establish the relative advantages of the proposed framework.

  2. Unclear independence of final evaluation from evolutionary search. Program structures are optimized using all behavioral samples, while validation scores and error cases repeatedly guide program modifications. The paper does not clearly establish whether the final results are evaluated on new trajectories or mazes that are completely independent of the evolutionary search process. This limits the assessment of the synthesized programs' generalization ability.

  3. Context-level KL divergence may not adequately capture state-level behavioral fidelity. The paper evaluates behavioral fidelity using context-level KL divergence, which maps different states into discrete contexts and compares aggregated action distributions. While this abstraction addresses the difficulty of estimating action distributions at individual states, it may obscure prediction differences across specific states. For example, a program could select actions opposite to those of the target model in different states while still producing highly similar aggregated action distributions. Although the paper also reports next-action prediction accuracy, low context-level KL divergence alone does not establish equally high fidelity at the individual decision level. Therefore, claims regarding behavioral fidelity should be interpreted more cautiously.

Questions for authors
  1. Could the authors include comparisons with other interpretable behavioral modeling methods, such as decision trees, rule-learning methods, or alternative program-induction approaches, to further validate the advantages of the proposed framework?

  2. Are the final behavioral fitting results obtained from data that were completely excluded from the evolutionary search process? Additionally, is the four-fold cross-validation performed at the level of complete task trajectories or individual decision records?

  3. Although Figures 8 and 9 report next-action prediction accuracy, Section 4.2 shows that situational KL divergence improves in all 36 settings, while prediction accuracy improves in only 32. Could the authors further analyze this discrepancy and clarify whether the situational abstraction affects the paper's main behavioral interpretations?

1 comment
Area_Chair_YwbCOpenReview ↗

Feedback

Hi, the review deadline is approaching. Please take a look at the authors' rebuttal and post your feedback (and update your rating if appropriate) at your earliest convenience. Thank you! AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.