{"id":"b810de74-6cb4-4aa6-a9e6-da9cc65d5b85","arxiv_id":"2504.20628","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"People navigating structured mazes appear to use modular, program-like cognitive maps that reuse repeated fragments, and a model combining LLM program synthesis with fragment-based planning predicts their paths.","lead":"This paper tests whether people navigate by compressing spatial maps into reusable programs, and reports that people search structured mazes in a modular, fragment-by-fragment way. It introduces a model that uses an LLM to infer such programmatic maps and predicts human search paths better than standard planning models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The untested load-bearing claim is that GPT4 program synthesis plus Eq. 2 scoring recovers the same fragment decomposition humans use; absent that, the behavioral fit could come from the scoring function rather than from generative-program cognition.","rationale":"The most load-bearing point is not that the behavioral effect is absent or fabricated; modular search in repeated mazes is plausible and the self-reported strategies likely support it. The issue is the mapping from behavior to mechanism. The paper's model only speaks to human cognition if the program-synthesis step recovers fragments that correspond to mental chunks; the text itself labels this an assumption. Because the main dependent measure and the model comparison are computed relative to that decomposition, a failure of the assumption would leave a behavioral modularity effect that can be explained by any chunking or coverage heuristic, while the specific 'generative program' claim would be unsupported. The reader's weakest_assumption identified exactly this LLM-to-human-prior link, and I agree. I would not move the verdict because the authors can satisfy the condition with a small human segmentation validation; if that validation fails, the model comparison should be downgraded. The absence of simple heuristic baselines is a reinforcing consideration rather than the primary one, because even a validated decomposition would not by itself show that programs (rather than hierarchical chunking) are needed; that contrast is the paper's stated differentiator from prior hierarchical models and remains untested.","tokens_in":11051,"tokens_out":5266,"duration_ms":58619,"concrete_test":"Run a fragment-segmentation experiment: naive participants view the 20 experimental mazes and mark repeated building-block regions; then compare human segmentations with GMM's fragments using adjusted Rand index, versus control segmentations from random rectangular tilings and from a no-LLM symmetry-based detector. If GMM-human agreement does not exceed controls, the LLM/scoring pipeline is not validated as a window onto human priors and the model-prediction results should be reinterpreted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, that human planning in structured environments is predicted by generative map programs and cannot be explained by unstructured cognitive-constraint planning, stands or falls on the fragment decomposition produced by the Generative Map Module. The paper states this explicitly in the Background: LLM program synthesis is used 'on the assumption that by doing so, we can implicitly access human prior knowledge of the world embedded in language and code used in LLM training.' This assumption is load-bearing because the main behavioral measure, a 'modular path,' is defined relative to fragments (visiting all fragments in order, moving to the closest subsequent fragment), and the model comparison in Fig. 4 computes discriminating decisions only after those fragments are fixed. The scoring function in Eq. 2 contains two free weights (w1, w2) balancing reconstruction error against description length; no values or sensitivity analysis are reported. If GPT4's fragment proposals are shaped by the prompt's helper functions, admissible transformations (rotation, reflection), and MDL pressure rather than by human chunking, then GMP's predictive advantage could be an artifact of the experimenters' scoring choices, not evidence for program-like cognitive maps. The Discussion acknowledges unmodeled variability and non-uniform fragment priors but does not address this grounding problem. A second gap reinforces the concern: no simple heuristic baseline (e.g., information-gain or nearest-unvisited-region planning) is compared, so even with clean fragments the unique contribution of 'generative programs' is not isolated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that human planning in structured environments is supported by cognitive maps represented as generative programs, in which repeated spatial fragments are discovered and planned over separately, yielding modular and resource-efficient but globally suboptimal behavior. The authors introduce the Generative Modular Planning (GMP) model, which uses GPT4 program synthesis to recover fragment decompositions from maze layouts and plans within each fragment, and they test it against two existing maze-search models using a 30-participant behavioral experiment. They report that people's search paths are predominantly modular according to their definition and that GMP predicts behavior better than Expected Utility and Discounted Utility planning. The central claim, in the paper's own words, is that human planning in structured environments is predicted by the generative map model and cannot be explained by unstructured planning with cognitive constraints.","tokens_in":11363,"tokens_out":4615,"duration_ms":51140,"significance":"If the central claim were fully supported, the paper would make a substantial contribution by connecting program induction, LLM-based priors, hierarchical planning, and behavioral data in a single framework. The experimental design is thoughtful: the mazes were constructed so that the alternative models predict non-modular policies or are indifferent, and the conservative modularity definition works against the authors' hypothesis. The use of GPT4 program synthesis as a bridge to human spatial priors is a novel and potentially valuable methodological step. The paper also provides a concrete efficiency argument for why program-structured maps could be adaptive. However, the strong comparative claim is currently under-supported: no inferential statistics are reported for the model comparison, the free weights in the scoring function are not specified or fitted, the LLM-derived fragment decomposition is not validated against human chunking, and the comparison set lacks simple heuristic baselines. These issues are load-bearing because the behavioral metric is defined relative to the GMM's fragments.","major_comments":[{"comment":"The scoring function in Eq. (2) contains two free weights, w1 and w2, but the manuscript never states their values, how they were chosen, or whether the reported results are sensitive to them. Algorithm 1 also leaves the threshold t and the number of completions C unspecified. Because the score determines which fragments GMM treats as programmatic structure, and because modular paths and discriminating decisions are computed relative to those fragments, the downstream model comparison in Fig. 4 can be shaped by these choices. Please report the parameter values, the fitting procedure (if any), and a sensitivity analysis over w1, w2, t, and C.","section":"Generative Map Module, Eq. (2), and Algorithm 1"},{"comment":"The claim that GMP predicts people 'significantly better' than the optimal planner or the limited-horizon planner is made without reporting any test statistic, p-value, effect size, or confidence interval for the model comparison. Fig. 4c shows per-participant fractions with 95% CIs, but no aggregate statistical test is reported. The assertion that the design makes alternative models 'indifferent between modular and non-modular search' with probability 'at most 0.25' also needs a derivation or supporting analysis. Without these quantitative details, the paper's central comparative claim is not established.","section":"Results, Fig. 4"},{"comment":"The paper explicitly relies on the assumption that GPT4 program synthesis 'can implicitly access human prior knowledge of the world embedded in language and code used in LLM training.' This assumption is load-bearing because the behavioral modularity metric is defined relative to the GMM's fragment decomposition: if GPT4's fragments do not correspond to the chunks humans actually use, the behavioral fit could come from the scoring function and prompt design rather than from generative-program cognition. The Discussion does not address this grounding problem. Please validate the fragment decomposition directly (for example, by eliciting human segmentations of the same mazes and measuring agreement with GMM fragments), and include a non-LLM baseline segmentation (such as wall-connected components, uniform tile grids, or an information-gain-based exploration policy) to test whether the predictive advantage comes from generative-program structure specifically.","section":"Background and Results"},{"comment":"The model comparison in the Alternative models section includes only the Expected Utility and Discounted Utility planners. The paper's strong claim that modular planning 'can not be explained by alternative models based on unstructured planning with cognitive constraints' is not supported by this comparison alone, because a simple greedy or nearest-unvisited-cell policy may also produce fragment-ordered search behavior in these mazes. Please add one or more simple heuristic baselines or explicitly weaken the claim to the two models actually tested.","section":"Alternative models"}],"minor_comments":[{"comment":"The exclusion of one maze because the exit was 'randomly placed in plain sight' is a post hoc exclusion and should be reported with a robustness check; if the maze is included in the analysis, the modularity results should be shown to be unchanged.","section":"Methods, Procedure"},{"comment":"The figure caption says the examples have 3, 2, and 5 structural fragments, while the text says 4, 2, and 5; this inconsistency should be corrected.","section":"Fig. 3"},{"comment":"There are several grammatical errors that should be fixed, including 'Our models leverages' in the abstract, 'We makes three contributions' in the Discussion, 'In the the remainder' near Eq. (1), 'perviously' in the Results, and 'Massachussetts' in the author affiliations.","section":"Abstract and Introduction"},{"comment":"The CRT-modularity regression is reported only as non-significant; please provide the correlation coefficient, sample size, and p-value so the reader can assess the null result, and consider whether the pilot-based sample size is sufficient for this secondary analysis.","section":"Discussion"},{"comment":"The manuscript does not include a data or code availability statement. For reproducibility, please provide the code for GMM/FP, the exact prompts used, and the parameter values for Algorithm 1 and Eq. (2).","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's interdisciplinary scope and the central idea is intriguing, but the strong comparative claim in the abstract is not yet supported by the reported evidence. The requested statistical reporting, parameter disclosure, and fragment-decomposition validation are necessary before I could recommend acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this paper has a genuinely new behavioral result and a model that is worth taking seriously, but the central mechanism — LLM program synthesis recovering human fragment decompositions — is under-specified, and the model comparison is missing the statistics needed to verify it.\n\nThe good parts first. The maze-search experiment is clean: 30 participants, structured mazes with repeated fragments, and the authors designed the environments so that the optimal planner and the discounted-utility model either predict non-modular paths or are indifferent. The conservative definition of a modular path (visiting fragments in order, moving to the closest next fragment) means the observed modularity is not inflated. The pilot-driven sample size is defensible. The paper also openly states the key assumption: GPT4 is used to implicitly access human priors. That honesty is a point in its favor.\n\nThe soft spots. The biggest one is the grounding of the fragment decomposition. The GMM uses GPT4 plus Eq.2 to find fragments, and the planning module then compares discriminating decisions with those fragments fixed. If GPT4's fragments come from prompt design or MDL pressure rather than human chunking, the behavioral fit could be a property of the scoring function, not of generative-program cognition. The free weights w1,w2 are never reported, and there's no sensitivity analysis. That is a real gap. Second, the model comparison in Fig.4 reports no statistical test values — just bars and confidence intervals. It's impossible to know whether GMP's advantage is significant. Third, there are no simple heuristic baselines (nearest-unvisited-region, information-gain) to rule out modularity without programs. Fourth, the abstract's claims about computational efficiency and memory are not measured anywhere in the paper. Minor point: the exclusion of the maze where the exit was in plain sight is fine — that's a procedural exclusion, not a results-based one.\n\nThere's also inherent circularity in designing the scoring metric around the modularity hypothesis, but the independent behavioral measurements mitigate that. The paper deserves a serious referee. I'd send it out with a request for code/data, w1,w2 values, sensitivity, baselines, and proper statistics.","headline":"A real behavioral effect with a plausible model, but the fragment inference step is under-specified and the model comparison lacks statistical detail.","tokens_in":11844,"tokens_out":3160,"would_cite":true,"duration_ms":31802,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that human planning in structured spaces works by representing maps as generative programs and reusing per-fragment plans, which makes search locally optimal and globally suboptimal, resource-efficient, and not…","keywords":["navigation","planning","cognitive maps","computational modeling","large language models","program induction","modularity","partial observability"],"falsifier":"Have participants directly segment the same mazes—mark the repeated chunks they perceive as units—and compare those chunks with the fragments GPT-4 recovers under Equation 2; if people segment the mazes differently on the very maps where the model predicts their search paths, the central claim that human plans track program-like fragments loses its support.","tokens_in":10876,"feed_emoji":"🧭","tokens_out":10608,"duration_ms":94742,"temperature":0.7,"pith_summary":"The paper is trying to establish that human navigation in structured, partially visible spaces is not driven by a full internal map or by shallow planning alone: people effectively store the environment as a generative program that describes repeated map fragments, plan a search policy once per fragment, and reuse that policy whenever the fragment reappears. This would explain why people are efficient in structured real-world environments yet look suboptimal under classical planning models: their plans are locally optimal inside each repeated unit and only globally suboptimal across fragments. To test this, the authors ran a maze-search experiment in which nearly all participants' paths showed a modular signature—visiting repeated fragments in order—across all analyzed mazes, and they built a computational model that recovers fragment programs using large-language-model program synthesis and plans fragment by fragment. The model predicted people's choices better than an optimal planner and a limited-horizon planner, while requiring far less computation and memory. The authors also report that a pilot correlation between cognitive-reflection scores and modularity did not replicate in the final sample.","feed_headline":"People navigate mazes by reusing repeated map fragments","feed_subtitle":"A generative-program model of cognitive maps predicts human search paths better than optimal or bounded planners.","key_machinery":"The load-bearing object is the generative map program: a compact program, synthesized by prompting a large language model, that reconstructs the observed grid from a small set of fragment units placed by translations, rotations, and reflections. Candidate programs are ranked by Equation 2, a weighted score combining grid-level reconstruction error with a minimum-description-length penalty on fragment complexity and transform counts. The Fragment-based Planning module then solves the partially observable Markov decision process (POMDP) belief-tree search once per fragment and reuses that decision subtree whenever the fragment recurs, using value iteration to move between fragments and falling back to non-modular planning when observations conflict with the reconstructed map. This machinery converts repeated structure into reusable computations, producing the predicted local-optimal/global-suboptimal search signature.","core_discovery":"The central claim, stated on the paper's own terms, is that human planning in structured environments is predicted by a generative-map model and cannot be explained by unstructured planning with cognitive constraints. A person's cognitive map is not a pixel-by-pixel layout but a small program that regenerates the environment from repeated fragments; the planner computes a policy once for each fragment and reuses it, and interpolates between fragments by solving a local search. Behavioral results show that people choose modular paths that visit all fragments in order in every analyzed maze, matching this signature of locally optimal, globally suboptimal planning, while the optimal expected-utility planner and the discounted-utility planning-horizon model predict non-modular or indifferent paths. The paper presents this as evidence that human deviations from optimality in structured environments arise, at least in part, from reasoning about environmental structure rather than only from limited planning depth.","pith_inferences":["Beyond the paper: if fragment policies are genuinely reusable, a person who has searched one maze containing a given fragment should navigate a second maze containing the same fragment faster, without relearning; this transfer effect is not tested in the paper and would be a direct behavioral check of the reuse mechanism.","Beyond the paper: the weighting parameters $w_1, w_2$ and the fragment priors could be fitted per participant, turning the paper's acknowledged variability in modularity into a measurement of individual differences in chunking rather than residual noise.","Beyond the paper: an independent probe of human chunking, in which participants freely segment maps into repeated units, could be compared against the fragments the language model recovers; agreement on the same mazes would separate the cognitive-claim contribution from the particular inductive biases of the language model."],"forward_implications":["In structured environments, human planning cost should scale with the number and complexity of distinct fragments rather than with map size, so large regular spaces should be much easier for people than their classical planning complexity suggests.","Planning models that ignore environmental structure, such as optimal belief-space search or a fixed discount factor, will systematically mispredict behavior in structured mazes even when they fit behavior in unstructured laboratory tasks.","LLM program synthesis can serve as a practical source of human-like priors about spatial chunking, since the recovered fragments are what let the model predict human search better than unstructured planners.","The remaining variability in behavior, with people sometimes skipping fragments, points to non-uniform beliefs about which fragments are likely to contain the reward; the paper identifies this as the next extension of the model."],"supporting_citations":[{"why":"Supplies the Maze Search Task, the optimal Expected Utility planner, and the Discounted Utility limited-horizon baseline that the generative model is compared against.","marker":"[Kryven et al., 2024]"},{"why":"Provides the POMDP belief-space tree search whose Bellman recursion underlies both the optimal planner and fragment-based planning.","marker":"[Kaelbling et al., 1998]"},{"why":"Establishes the intractability of probabilistic planning, motivating the need for structured programmatic approximations.","marker":"[Madani et al., 2003]"},{"why":"Provides evidence that people construct simplified mental representations to plan, the premise that programmatic maps formalize.","marker":"[Ho et al., 2022]"},{"why":"Supports the concepts-as-programs view that the paper extends to cognitive maps.","marker":"[Lake and Piantadosi, 2020]"},{"why":"Supplies the minimum-description-length principle behind the fragment scoring function in Equation 2.","marker":"[Rissanen, 1978]"},{"why":"Demonstrates using LLM program synthesis to build world models, the method the Generative Map Module adapts.","marker":"[Tang et al., 2024]"},{"why":"Provides prior work on compositional spatial submap learning that the generative-map approach extends with LLM synthesis.","marker":"[Sharma et al., 2022]"}],"fun_headline_variants":["Minds see space as programs, not pixels","Planning works by reusing map pieces, not full maps","The brain's map is a compact program","Humans plan with map fragments, not whole layouts","Why people take suboptimal paths: they use modular maps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the fragment decomposition recovered by GPT-4 program synthesis under the paper's scoring function matches the fragments people actually form in their heads; the authors state this as an assumption that LLM training data gives implicit access to human prior knowledge, and if that link fails, the behavioral fit could be an artifact of scoring weights and prompt design rather than evidence about human cognition.","fun_headline_variants_meta":{"raw":{"variants":["Minds see space as programs, not pixels","Planning works by reusing map pieces, not full maps","The brain's map is a compact program","Humans plan with map fragments, not whole layouts","Why people take suboptimal paths: they use modular maps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000997,"raw_usage":{"total_tokens":4215,"prompt_tokens":933,"completion_tokens":3282,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":3207}},"tokens_in":549,"tokens_out":3282,"duration_ms":22968,"temperature":1.0,"reasoning_tokens":3207,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:24:30.718631+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have participants directly segment the same mazes—mark the repeated chunks they perceive as units—and compare those chunks with the fragments GPT-4 recovers under Equation 2; if people segment the mazes differently on the very maps where the model predicts their search paths, the central claim that human plans track program-like fragments loses its support.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Maze Search Task, the optimal Expected Utility planner, and the Discounted Utility limited-horizon baseline that the generative model is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the intractability of probabilistic planning, motivating the need for structured programmatic approximations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the concepts-as-programs view that the paper extends to cognitive maps."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides prior work on compositional spatial submap learning that the generative-map approach extends with LLM synthesis."}],"review_version":1}