acceptodds
Under review as a conference paper at ICLR 2027

Generative Cognitive Maps for Structure-Sensitive Planning

Abstract

Existing models of human planning generally explain human behavior through limited planning depth or resource-rational constraints on computation. While such models are agnostic to environmental structure, real-world environments such as natural landscapes and building interiors often contain repeated or compositional structure. Here we test experimentally whether human planning exploits such structure, finding strong structure sensitivity that is not captured by current state-of-the-art models of human planning. To instantiate this principle computationally, we propose GenPlan, a computational framework that shows how structure sensitivity can be implemented in practice through compositional generative map representations and reusable policies. The framework integrates (1) a Generative Map Module that infers latent structural regularities, and (2) a Structure-Based Planner that exploits them for hierarchical planning. We show that GenPlan captures sensitivity to structure empirically observed in human behavior, while reducing planning cost relative to structure-agnostic baselines with only a bounded performance penalty. Together, these results establish structural sensitivity as an important principle of human planning and demonstrate how it can be incorporated into resource-efficient planning algorithms.

Primary area:
applications to neuroscience & cognitive science
Keywords:
cognitive maps, human planning, compositionality, program induction, structure-sensitive planning

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

This paper examines how repeated spatial structure affects human decision-making in a maze navigation task with partial observability, postulating that human search in a structured environment is not approximated by local search methods.

The authors propose the GenPlan framework, in which an LLM first derives a programmatic decomposition of a map into repeated units, inspired by compositional accounts of human cognition, which is then passed to a structure-based planner combining POMCP-based search within repeated units with between-unit transitions.

Experiments aim to verify how closely navigation approaches used by 30 human crowd workers match planning policies with (Gen-POMCP) and without (DepthPlan, Naive-POMCP) map decompositions into structural sub-units, focusing in particular on discriminating decision points where planning policies diverge from human behavior.

Results show that Gen-POMCP outperforms both DepthPlan and Naive-POMCP in predicting recruited participants' behavior. Simulations on ten larger mazes show better exploration per rollout, with performance regressing toward Naive-POMCP as the proportion of unobserved structured units decreases with exploration.

Critical strengths
  • The proposed GenPlan approach is well-motivated by observations of compositional structures in human cognition, and outperforms the reported baselines in predicting actual human behavior from paid crowd workers in the evaluated MST setting (80.7% vs 42.2% for DepthPlan and 42.5% for Naive-POMCP). These results provide evidence that explicit structural biases can better approximate human search decisions.

  • The additional ablation on an end-to-end LLM-based planner without program induction (LLM-Direct, Appendix I) shows that a naive LLM use is not sufficient for robust navigation, motivating the use of explicit planning modules adopted in the main experimental section.

Critical weaknesses
  • The central behavioral comparison remains insufficiently specified. Because agreement with human behavior is the core evaluation metric, the discussion should clearly present disagreements between human participants to contextualize results against baseline population variance. This under-specification also appears in Appendix I, where it is unclear whether the weak performance observed across several LLMs and reasoning budgets can be attributed to a suboptimal problem formulation. While Appendix E shows examples of decomposed units from the map, the paper does not include an explicit ablation study showing how different decomposition quality affects the structure-based planner's performance, particularly for the less well-defined policy ruling transitions between repeated units. Without these clarifications, it is difficult to assess whether the authors' comparison is meaningful and reproducible.

  • The proposed Gen-POMCP strategy proves especially effective only for partially observable navigation scenarios in environments with known structural boundaries and repeated structural units. As the authors note in Section 3.2, their method reverts to naive global planning when such regular structures are absent. While the results seem promising in the handcrafted MST settings the authors study, I believe this setup is too narrow to provide convincing evidence that LLM-driven decomposition strategies could be applicable to more realistic planning scenarios.

Questions for authors
  • Could you clarify the prompting setup that was used for the LLM-Direct evaluation? In particular, how was the current observation sent to the model? Were models asked to take one action at a time to account for updated observations, or were they planning entire trajectories from their initial observations?
Reviewer_NdP9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether humans exploit repeated environmental structure to simplify planning. It proposes GenPlan, which uses an LLM to infer recurring spatial units and generate map reconstruction programs, then computes and reuses local search policies through a conventional planner. Human maze experiments show higher agreement with participants’ choices than several structure-agnostic alternatives, while simulations demonstrate reduced search computation. The paper also provides a conditional performance bound for an ideal modular-optimal policy.

Critical strengths

S1. A meaningful research question with substantial aggregate behavioral differences The paper directly tests competing explanations of human planning, including global planning, limited planning depth, and nearest-first search. The reported agreement across all decisions is 80.7% for GenPlan, compared with 42.5% for Naive-POMCP and 42.2% for DepthPlan. The additional comparison with the Steps heuristic further supports the narrower conclusion that these alternatives do not adequately explain behavior in the tested structured environments.

S2. A clear connection between structural inference and policy reuse, supported by targeted evaluations GenPlan connects inferred map decompositions to reusable local policies conditioned on unit entrances. With the same POMCP backend, it requires substantially fewer search rollouts, and its advantage decreases as reusable structure is progressively disrupted. These evaluations support the computational value of structural reuse, while the theoretical analysis provides a conditional bound for optimal planning under the modular constraint.

Critical weaknesses

W1. The behavioral results conflate individual and group statistics, with direct inconsistencies between the text and figures The main text states that every participant’s modular fraction exceeds 80% and that all individual confidence intervals exclude 0.5. However, Figures 4b and 13 include individual results near 0.5, while Figure 12 presents only two estimates aggregated across participants. These discrepancies do not directly invalidate the reported aggregate advantage, but they undermine claims of universal individual support. The authors should reconcile the figures and text and clarify the denominators, aggregation procedures, and confidence interval calculations.

W2. The trade-off between total computational cost and path quality is insufficiently validated, and the implemented planning objective remains unclear The simulations primarily measure full-map exploration against rollout counts, without jointly evaluating expected steps to reach a uniformly sampled exit and total computational cost, including GMM. The theoretical guarantee applies only to an ideal modular-optimal policy. Moreover, the reward for reaching the exit in the main text is not clearly reconciled with termination upon observing the exit in the pseudocode, leaving the implemented objective and its relationship to the theoretical analysis insufficiently specified.

Questions for authors

Q1. What aggregate effect and individual variation remain after correcting the statistical reporting? Please provide participant-level counts of discriminating decisions, modular fractions, and confidence intervals, explain the aggregation procedure, and reconcile the inconsistencies between the main text and Figures 4b, 12, and 13.

Q2. Does GenPlan retain an efficiency advantage under a clearly specified objective and complete cost accounting? Please clarify how rewards and termination distinguish observing from reaching the exit, and report the trade-off between total runtime—including GMM—and expected steps to reach a uniformly sampled exit.

Reviewer_x8qJOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The main result of the paper is to show that humans exploit repetition in structure when planning. The authors use the Maze Search Task (MST) to exemplify this. They compare four different computational models for planning: three of them do not exploit structure repetition and one of them (Gen-POMCP) does so by inferring a distribution p(M∣I)p(\mathcal{M}|I) where M\mathcal{M} is a partition of repeating structural units in the map. They then run a human experiment on various MST maps and show that Gen-POMCP aligns mostly closely to the set of actions human take when exploring the map.

Critical strengths

There are no critical strengths, but I do want to highlight two positive aspects of the paper:

  • I appreciate the level of detail that went into the construction of Gen-POMCP.
  • I thought LLM-Plan was an interesting baseline to include and feels like an interesting direction to explore. E.g. authors note that "strongest condition tested — GPT-5.2 at high reasoning effort on small maps — produced valid paths but substituted a full-coverage objective for the correct expected-cost objective, visiting every cell rather than planning against a uniform prior over exit locations", but what if I sharpened my prompt to include some form of resource constraint?
Critical weaknesses
  1. Significance: A priori, it is not clear to me why the main result of the paper is especially surprising or interesting. The fact that humans do not plan optimally is a well-known result. As the authors themselves note, prior explanations generally revolve around some form of resource constraint, i.e. limited depth planning. However, exploiting repetition in structure is itself a way to plan more efficiently, by allowing humans to transfer knowledge from previous exploration without additional cognitive resources. In fact, the authors themselves demonstrate this in Section 3.2, by showing that Gen-POMCP requires a smaller compute budget than Naive-POMCP. The reality is probably that neither simplistic explanation (limited depth or structure exploitation) can fully account for human planning and the kind of cognitive biases that dominate are likely to be highly problem-dependent and resource-dependent.

  2. Methodology: This is related to my first point. Because the authors chose small MST maps with highly repetitive structure, limited depth, and compared only to naive/non-human methods (Naive-POMCP and LLM-Plan) and a single other cognitive model (DepthPlan), it is not at all surprising that Gen-POMCP aligns mostly closely to the very limited (!) (N=30) samples of human trials here. In some sense it almost feels like the problem was designed to fit the solution. Two of the three baselines are already unrealistic, and more importantly, Gen-POMCP learns an extremely informative posterior because the environment itself is easily compressible. See Question 2 below.

Questions for authors
  1. Can you design maps with no repeated structures at all? In these cases, does Gen-POMCP's solution converge to Naive-POMCP's? Do they also converge in terms of compute budget?

  2. The criticism that limited depth planning diverges from human behavior feels unfair for the small maps used in the paper (Appendix C). For such small maps where the structure repetition is obvious even early on in exploring, I can't imagine depth constraint to be a real issue here. E.g. looking at Figure 3b, there is no reason why I might want to take the straight path when the map is small enough that I can afford to start exploring the first subsection at depth.

What happens when I scale the maps to far larger sizes than those used in the paper, with some degree of repetitive structures spaced far apart? Is Gen-POMCP still an accurate model?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.