acceptodds
Under review as a conference paper at ICLR 2027

Agent-Based Human Mobility Simulation with Personalized Cognitive Maps

Abstract

Large language models enable agents to simulate human mobility through sequential decisions about daily activities and destinations, without requiring task-specific training on large-scale trajectory data. However, existing methods often overlook how each individual’s limited and evolving knowledge of the city constrains their mobility decisions. To address this limitation, we introduce CogMob, which equips each agent with a personalized, partial, and evolving cognitive map and incorporates it into mobility decision-making, where intent-aware retrieval activates relevant spatial knowledge and destination selection uses this knowledge to compare candidates in light of the agent’s accumulated experience and current context. Experiments on two public datasets show that CogMob consistently improves mobility simulation across a wide range of metrics. We further assess its downstream utility through segregation analysis, observing patterns qualitatively consistent with prior findings. Our code is available at https://anonymous.4open.science/r/CogMob-17B2.

Primary area:
generative models
Keywords:
Human mobility simulation, Cognitive maps, Large language models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_QESNOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CogMob, an LLM-based agent framework for simulating individual daily mobility without training on trajectory data. Each agent keeps a personal cognitive map of the places it has visited or heard of, its routes and the regions it knows. The map is updated through experience and slowly forgets places that are not revisited. Candidate destinations are retrieved from the map (plus new POIs from a database) according to the agent's spatial intent, and the LLM picks one. Experiments on Beijing and NYC compare CogMob with training-free LLM simulators and trained generative models, with ablations, a plug-in experiment with LLMob and a segregation analysis.

Critical strengths

Current LLM mobility simulators do not reproduce how differently individuals explore new places and return to familiar ones (Fig. 2), and giving each agent its own evolving map of what it knows is a natural way to address this. The ablations support it. Removing the map clearly hurts the exploration and return metrics (Table 3), and adding it to LLMob helps as well (Fig. 5). I find this part convincing, and it's the main reason I'm positive about the paper.

The evaluation also goes further than most papers in this line, with metrics that test known mobility regularities and data-fitted baselines such as TimeGeo.

Critical weaknesses
  1. My main concern is that Table 1 overstates what the cognitive map itself contributes. On Beijing, the variant without the map (w/o CM-B) already beats the training-free baselines on most metrics, so much of the gain seems to come from the rest of the agent design, some of which the adapted baselines lack. The map clearly matters for exploration and return, and I would frame the claims around that.

  2. It is not clear what the map's structure adds over simply remembering visited places. The visit-history baseline in App. D.4 looks like a separate, simpler simulator rather than CogMob with a flat memory.

  3. No runtime or cost is reported. Each decision needs at least two LLM calls plus periodic reflection, and App. B.4 runs at most five agents at a time, so it is hard to judge how this would scale to realistic population sizes.

  4. There are many hand-set parameters (decay rates, search radii, the clustering radius, candidate limits), and I could not find a sensitivity analysis or an explanation of how they were chosen.

Questions for authors
  1. Could you run CogMob with the cognitive map replaced by a flat list of visited places, keeping everything else the same? If the structured map still clearly helps on the dynamics metrics, that would address my main concern.
  2. How does the cognitive map relate to the per-agent environment trees in Generative Agents (Park et al., 2023), which also start partial, grow with exploration and are used to choose locations?
  3. For NYC, how are the simulated full-day trajectories made comparable to the sparse check-in data before computing the JSDs? Some of the NYC differences also look small relative to the run-to-run variation in Table 4.
  4. The segregation analysis doesn't seem to separate models much (CoPB also beats the shuffled baseline, Fig. 9), and for Beijing the reference data come from US cities. Unless I'm missing something, I would present it as a sanity check rather than as a contribution.
8 comments
AuthorsOpenReview ↗

Response to Reviewer QESN (1/6)

Q1: Could you run CogMob with the cognitive map replaced by a flat list of visited places, keeping everything else the same? If the structured map still clearly helps on the dynamics metrics, that would address my main concern.

Response: We thank the reviewer for this question. We have added an ablation variant, CogMob with Visited-only Flat Memory (CogMob w VFM), which replaces the cognitive map with a flat list of visited places while keeping the remaining components unchanged. Given the intent generated in Stage 1, Stage 2 retrieves candidates, ranked by visit frequency and distance. If fewer than nine candidates are available, it supplements them through a search of the full POI database.

As shown in the table, CogMob reduces JSD by an average of 31.52% compared with this variant. Flat visitation memory records where an agent has been, but does not fully capture its personal understanding of the city: which places it knows beyond those already visited, how places are connected through routes and organized into regions, and how this knowledge evolves over time. CogMob explicitly models these aspects and uses them to guide destination selection.

DatasetMethodRadiusJumpDurationEntropyExploreReturnVisit-FreqCapacityLocFreqOdSim
CogMob w VFM0.03140.03950.04120.15920.03980.09940.16260.00080.18220.3294
CogMob0.02140.02920.03780.13310.02710.02950.09110.00030.15440.2943
Improvement31.85%26.08%8.25%16.39%31.92%70.32%43.97%60.47%15.26%10.66%
AuthorsOpenReview ↗

Response to Reviewer QESN (2/6)

Q2: How does the cognitive map relate to the per-agent environment trees in Generative Agents (Park et al., 2023), which also start partial, grow with exploration and are used to choose locations?

Response: We thank the reviewer for this question. In terms of purpose, Park et al.’s environment tree helps agents choose where to carry out activities in a predefined town with only 19 top-level venues, with several categories represented by a single venue, such as a café, pub, or park. CogMob addresses how individuals choose among the many destinations available for the same activity in a city, and why they repeatedly visit only a small, changing subset of these alternatives. Its cognitive map models how spatial knowledge and experience mediate this selective use of urban space, allowing personal activity spaces to form and evolve through successive destination choices.

Methodologically, the two approaches differ in how spatial knowledge is formed and updated, and how it supports decisions at different spatial scales. In Park et al.’s approach, each agent starts with partial knowledge of a predefined “world–venue–subarea–object” containment hierarchy and expands it through direct observation. Given an intended activity, the LLM selects a known venue, then a subarea and a specific object within it. However, simply accumulating observations alone does not fully capture how individuals understand and use urban space. Even people who visit the same places may form different impressions, remember different nearby POIs, perceive accessibility differently, and group places into different personal regions. These differences shape which destinations they consider, revisit, or explore.

In contrast, CogMob organizes spatial knowledge into places, routes, and regions. It distinguishes visited places from those the agent only knows about and forms personal understanding of regions by clustering visited places. This knowledge evolves with new information exposure and new experience. Agents learn about unvisited places through indirect information relevant to their interests and current intent. Reflection revises personal impressions of visited places, and memory strength decays over time. These updates change what the agent remembers, how it perceives places, and how its spatial knowledge is organized. CogMob then uses this evolving knowledge to support destination choice among urban POIs. Depending on whether the agent has a specific place, a target region, or only preferences in mind, intent-guided retrieval generates destination candidates. This process links what an individual knows and remembers about the city to the destinations they consider visiting.

Experimentally, to compare CogMob with the environment-tree mechanism, we introduce a variant, CogMob with Perception-expanded Flat Memory (CogMob w PFM), which replaces the cognitive map with a place memory. After each visit, this memory adds the visited place and all nearby POIs, following the observation-based expansion used in Park et al.’s environment trees. All other components and experimental settings remain unchanged. As shown in the table, CogMob achieves an average improvement of 33.36% over this variant on the Beijing dataset.

DatasetMethodRadiusJumpDurationEntropyExploreReturnVisit-FreqCapacityLocFreqOdSim
BeijingCogMob w PFM0.03250.03700.03730.14820.05430.09070.20280.00090.18780.3435
CogMob0.02140.02920.03780.13310.02710.02950.09110.00030.15440.2943
Improvement34.15%21.08%-1.34%10.19%50.07%67.47%55.07%64.79%17.78%14.32%
AuthorsOpenReview ↗

Response to Reviewer QESN (3/6)

Q3.A: For NYC, how are the simulated full-day trajectories made comparable to the sparse check-in data before computing the JSDs?

Response: We thank the reviewer for this question. Check-in data are widely used to evaluate human mobility simulations in prior studies such as LLMob, CoPB, and MIRAGE. Despite being sparse compared to dense data like GPS trajectories, check-in data are favored considering its public availability and effectiveness in representing human mobility. Our evaluation compares the distributions of mobility statistics that characterize human mobility patterns, rather than individual full-day trajectories. Before using the NYC POI check-in data for evaluation, we verified that the patterns considered in our study remained observable despite the sparsity of the records (Figures 14–19). We then computed the corresponding statistics from the simulated trajectories and used JSD to assess how closely the simulated distributions match those observed in the check-in data.


Q3.B: Some of the NYC differences also look small relative to the run-to-run variation in Table 4.

Response: We thank the reviewer for this comment. Each experiment was repeated three times, and Table 4 reports the mean and standard deviation across runs. Among the methods listed, CogMob achieves the lowest mean JSD on 12 of the 13 NYC metrics. To examine the smaller margins relative to run-to-run variability, we inspected the results for user-level Explore, Return curve, Capacity, and LocFreq. For the first three metrics, CogMob achieved a lower JSD than the strongest baseline in all three runs, with relative reductions in mean JSD of 7.81%, 19.35%, and 11.54%, respectively. Thus, although the absolute differences are small, the improvements were consistent across the evaluated runs.

The advantage on LocFreq is more limited: CogMob outperformed the strongest baseline in two of the three runs, with a mean JSD reduction of 3.34%. This metric measures aggregate spatial visitation shares, so better individual exploration and return behavior need not produce an equally large improvement in LocFreq. NYC’s smaller study area may also allow distance-based selection to capture the main activity areas reasonably well, leaving less room for improvement. By comparison, cognitive-map-based retrieval may offer greater benefits in the larger Beijing study area by identifying relevant destinations outside agents’ immediate neighborhoods.

Reviewer_pc7JOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CogMob, an LLM-based framework for agent-driven human mobility simulation incorporating personalized cognitive maps. To model individuals' partial and evolving knowledge of urban environments, CogMob maintains a hierarchical spatial memory consisting of places, routes, and regions. The framework dynamically updates this memory through simulated experiences, indirect exposure, and reflection, while employing intent-guided retrieval to support personalized destination selection. Experiments on Beijing and NYC datasets demonstrate improvements over several training-free and training-based mobility simulation baselines across multiple metrics. The authors further investigate the utility of simulated trajectories through socioeconomic segregation analysis.

Critical strengths
  1. The paper addresses a meaningful problem by incorporating individualized spatial knowledge into human mobility simulation.

  2. The framework is clearly presented and achieves encouraging empirical improvements across multiple mobility metrics.

Critical weaknesses
  1. Insufficient methodological novelty and differentiation from prior work.
    The proposed framework mainly combines hierarchical spatial memory, heuristic memory updating and forgetting, and LLM-based retrieval and decision-making. These components are relatively conventional, and their integration does not appear to introduce substantial methodological innovation. In particular, the paper does not sufficiently establish how its cognitive-map design fundamentally differs from existing memory-augmented mobility agents such as GATSim, AgentMove, ZHMF, and CitySim. A more thorough comparison is needed to clarify the actual novelty of the proposed approach.

  2. Insufficient empirical evidence to support the central claims.
    Although the ablation studies show performance gains, they do not convincingly isolate the contribution of the proposed hierarchical cognitive maps from those of general personalized memory and improved candidate retrieval. The simple visitation-history baseline is insufficient; stronger memory-based baselines and component-level ablations are needed to establish whether the cognitive-map structure is genuinely necessary. Moreover, the main experiments are restricted to two cities and one-month simulation periods, leaving the robustness of the framework across longer simulation horizons, diverse populations, and different urban environments underexplored. Improvements in distributional similarity also do not necessarily establish that the simulated agents possess realistic human spatial cognition.

Questions for authors
  1. What are the key methodological differences between CogMob and existing personalized memory frameworks, particularly GATSim, CitySim, and AgentMove? Can the authors provide more convincing evidence of the benefits specific to cognitive maps?

  2. How sensitive is the framework to the design of its memory hierarchy, initialization, and forgetting mechanisms? Can the authors provide stronger generalization evidence across different temporal or geographic settings?

3 comments
AuthorsOpenReview ↗

Response to Reviewer pc7J (1/3)

Q1: What are the key methodological differences between CogMob and existing personalized memory frameworks, particularly GATSim, CitySim, and AgentMove? Can the authors provide more convincing evidence of the benefits specific to cognitive maps?

Response: We thank the reviewer for this question. ZHMF and AgentMove use memory to support mobility prediction, whereas CitySim and GATSim use memory to support mobility simulation. CogMob differs from these frameworks by modeling an evolving, subjective representation of the city that captures how individuals remember, perceive, and spatially organize their surroundings, and how this understanding guides mobility decisions beyond past visit patterns. We further assess the potential of our synthetic mobility data for socioeconomic research through segregation analysis, an application not examined in these studies. The results reveal patterns qualitatively consistent with prior empirical findings by Moro et al. [6], which cannot be achieved using the baseline method (as shown by the comparison with CoPB in Appendix C.4). Below, we clarify the methodological differences and present ablation evidence isolating the benefits of cognitive maps.

For methodological differences:

  • Mobility prediction: ZHMF stores past visits and LLM-generated reflections, while AgentMove maintains mobility statistics and recent visit information. These memory designs primarily capture past mobility behavior rather than provide a structured representation of individual spatial knowledge. Specifically, they do not explicitly represent known but unvisited places, how individuals organize space into regions, or their subjective perceptions of connections and distances between places and regions. Such spatial knowledge also guides exploration. Although AgentMove draws on the LLM’s geographic knowledge to suggest unvisited destinations, these suggestions serve as prediction-time cues rather than persistent additions to individual memory. It therefore expands the candidate set without modeling which unvisited places an individual knows or how that knowledge accumulates.
  • Mobility simulation: GATSim primarily maintains episodic memory rather than spatial memory, and its 13-location simulation setting provides limited evidence of scalability to city-scale environments. CitySim maintains memories of visited locations, including agents’ subjective impressions of those places. It shares the same limitation: its memory captures past experiences without explicitly modeling individuals’ broader spatial knowledge or subjective understanding of space. We also include CitySim as a baseline, and CogMob outperforms it in our experiments.

Experimentally, we conduct two ablation studies to compare cognitive maps with the two memory designs discussed above: episodic memory and flat memory of visited locations. (1) CogMob w/o CM-A and w/o CM-B (Section 4.4). This variant retains episodic memory but removes the cognitive map. Compared with w/o CM-B, CogMob improves eight of the ten metrics in Beijing, with an average relative reduction in JSD of 30.12% across all ten metrics. This suggests that the cognitive map contributes more than simply retaining past experiences. (2) CogMob with Visited-only Flat Memory (CogMob w VFM). This variant replaces the cognitive map with a flat memory of visited locations while keeping all other components unchanged.

Compared with this variant, CogMob reduces JSD by an average of 31.52%. Visited-only flat memory records where an agent has been, but does not fully capture its personal understanding of the city: which places it knows beyond those already visited, how places are connected through routes and organized into regions, and how this knowledge evolves over time. CogMob explicitly models these aspects and uses them to guide destination selection.

DatasetMethodRadiusJumpDurationEntropyExploreReturnVisit-FreqCapacityLocFreqOdSim
CogMob w VFM0.03140.03950.04120.15920.03980.09940.16260.00080.18220.3294
CogMob0.02140.02920.03780.13310.02710.02950.09110.00030.15440.2943
Improvement31.85%26.08%8.25%16.39%31.92%70.32%43.97%60.47%15.26%10.66%

[6] Moro, Esteban, et al. "Mobility patterns are associated with experienced income segregation in large US cities." Nature communications 12.1 (2021): 4633.

AuthorsOpenReview ↗

Response to Reviewer pc7J (2/3)

Q2.A: How sensitive is the framework to the design of its memory hierarchy, initialization, and forgetting mechanisms?

Response: We thank the reviewer for this comment. To examine how the framework depends on its memory representation, updating mechanism, and initialization, we evaluate the following variants on the Beijing dataset:

  • Memory hierarchy: CogMob with Visited-only Flat Memory (CogMob w VFM). This variant replaces the three-layer cognitive map with a flat memory containing only visited places. It retains the ability to accumulate visitation history but removes known-of places, route knowledge, and region-level organization.
  • Forgetting mechanisms: CogMob without forgetting (CogMob w/o forgetting). This variant retains the full model and all updating mechanisms, but disables forgetting components.
  • Initialization: CogMob without Initialization (CogMob w/o Init). This variant removes all initial non-anchor knowledge from the cognitive map, including visited places (L1a), known-of places (L1b), routes (L2), and regions (L3), retaining only home and workplace as spatial anchors. The cognitive map subsequently develops through the standard updating mechanism.

As shown in the table, CogMob achieves average improvements of 31.52%, 14.61%, and 20.30% across the ten metrics over CogMob w VFM, CogMob w/o forgetting, and CogMob w/o Init, respectively. These results show that the memory hierarchy, forgetting mechanism, and initialization each contribute to improved simulation performance.

DatasetMethodRadiusJumpDurationEntropyExploreReturnVisit-FreqCapacityLocFreqOdSim
BeijingCogMob w VFM0.03140.03950.04120.15920.03980.09940.16260.00080.18220.3294
CogMob w/o forgetting0.02200.03210.03970.13670.03840.04480.12980.00040.16590.3116
CogMob w/o Init0.02280.03790.04520.13470.02800.06480.14110.00060.16540.3151
CogMob0.02140.02920.03780.13310.02710.02950.09110.00030.15440.2943

Q2.B: Can the authors provide stronger generalization evidence across different temporal or geographic settings?

Response: We thank the reviewer for the suggestion. Our original experiments cover two cities and one month, a scope comparable to prior mobility simulation studies. These studies generally evaluate a limited number of cities, including CoPB (Beijing and Tokyo), LLMob (Tokyo and Osaka), CitySim (Tokyo), MIRAGE (Tokyo, Istanbul, and New York City), TimeGeo (Boston), MoveSim (Beijing), and Geo-Llama (Beijing plus a synthetic dataset), with evaluated generation horizons ranging from single-day trajectory generation to month-scale simulations.

Nevertheless, to strengthen the generalization evidence, we have added an experiment on the Tokyo Foursquare POI check-in dataset [1]. Following the same procedure used for NYC, we generate agent profiles based on census information. We simulate the mobility of 1,000 agents throughout June 2012 and evaluate the generated trajectories against all real check-in records observed during the same month. Because generating trajectories with LLM-based baselines takes considerable time, we selected CoPB, the best-performing LLM baseline on the Beijing and NYC datasets, for comparison within the rebuttal period. We also included MIRAGE and TimeGeo, two non-LLM methods that performed well on both datasets and require training.

As shown in the table, CogMob outperforms TimeGeo, MIRAGE, and CoPB on seven, nine, and eight of the ten metrics in Tokyo, respectively, achieving average relative improvements of 8.66%, 16.75%, and 32.38% across all ten metrics. These results provide further support for CogMob’s generalizability across different urban environments.

DatasetMethodRadiusJumpDurationEntropyExploreReturnVisit-FreqCapacityLocFreqOdSim
TokyoTimeGeo0.09970.01420.23300.31100.23900.21570.26130.00290.03440.2425
MIRAGE0.19530.19440.23320.10230.05600.14890.14550.00360.02570.4079
CoPB0.02350.01990.33570.26020.20040.30980.62680.00900.07660.3476
CogMob0.02320.04030.22000.09930.05240.06270.03580.00020.09290.3310

[1] Yang, Dingqi, et al. "Modeling user activity preference by leveraging user spatial temporal characteristics in LBSNs." IEEE Transactions on Systems, Man, and Cybernetics: Systems 45.1 (2014): 129-142.

AuthorsOpenReview ↗

Response to Reviewer pc7J (3/3)

W2: Improvements in distributional similarity also do not necessarily establish that the simulated agents possess realistic human spatial cognition.

Response: We thank the reviewer for this comment. Directly testing whether agents’ spatial representations and cognitive processes correspond to those of humans would require targeted human behavioral data, such as spatial judgments and planning-task responses, which are not available in our mobility datasets. Our goal is to improve mobility simulation by modeling how individuals’ spatial knowledge shapes their destination choices. Our modeling is motivated by evidence that human mobility decisions are influenced by individuals’ spatial knowledge and how that knowledge is organized and used [2-3]. Previous research shows that spatial knowledge is often incomplete [4] and can exhibit person-specific clustering and hierarchical organization [5]. These findings inform CogMob’s design: it represents individuals’ knowledge of the city through place, route, and region layers, updates this knowledge through individual experience, and retrieves relevant knowledge according to the current intent to guide destination selection. Our experiments assess whether incorporating these cognitively informed mechanisms improves mobility simulation.

Our evaluation also (1) examines whether the generated trajectories capture empirical behavioral relationships. Appendix E provides visual comparisons of observed and simulated exploration growth, preferential return, and the growth of activity-location sets. These comparisons consider both population-level regularities and variation across individuals. (2) The segregation analysis provides a further test because matching conventional mobility statistics does not necessarily imply that generated trajectories capture socioeconomic structure. For example, Moro et al. [6] show that the standard EPR model explains certain visitation patterns but does not capture observed differences in experienced income segregation. Our results show that the simulated data reproduce both non-random socioeconomic segregation and its relationship with exploration behavior. Specifically, individual experienced segregation and place segregation are higher than under the shuffled baseline, while greater place exploration and social exploration are generally associated with lower experienced segregation, consistent with empirical findings [6].

[2] Hannes, Els, Davy Janssens, and Geert Wets. "Destination choice in daily activity travel: mental map's repertoire." Transportation research record 2054.1 (2008): 20-27.

[3] Behrens, Timothy EJ, et al. "What is a cognitive map? Organizing knowledge for flexible behavior." Neuron 100.2 (2018): 490-509.

[4] Tversky, Barbara. "Cognitive maps, cognitive collages, and spatial mental models." European conference on spatial information theory. Berlin, Heidelberg: Springer Berlin Heidelberg, 1993.

[5] Hirtle, Stephen C., and John Jonides. "Evidence of hierarchies in cognitive maps." Memory & cognition 13.3 (1985): 208-217.

[6] Moro, Esteban, et al. "Mobility patterns are associated with experienced income segregation in large US cities." Nature communications 12.1 (2021): 4633.

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

This paper proposes CogMob, a cognitive map-guided framework for LLM-based human mobility simulation. The central idea is to equip each agent with a personalized, partial, and evolving cognitive map, allowing it to make mobility decisions based on its accumulated spatial knowledge rather than relying solely on static contextual information.

Specifically, CogMob introduces a three-layer cognitive map representation, together with mechanisms for dynamically updating spatial knowledge and retrieving relevant information according to an agent's current intentions. The framework aims to produce more realistic human mobility behavior by explicitly modeling how individuals perceive, remember, and interact with their surrounding environments.

Experiments on datasets from Beijing and New York City demonstrate improvements across multiple mobility simulation metrics. The authors further evaluate the practical utility of the simulated trajectories through segregation analysis, showing that the resulting patterns are qualitatively consistent with previous empirical findings.

Overall, I find the motivation insightful and the study reasonably complete. However, I have concerns regarding the limited diversity of evaluation datasets, the methodological novelty of the agent-based implementation, and the lack of comparisons with established trajectory generation models.

Critical strengths
  1. Insightful motivation supported by empirical observations. I find the motivation section particularly interesting. The authors investigate the importance of cognitive maps in human mobility simulation and provide empirical analyses to support their role in mobility decision-making. This offers a meaningful perspective on incorporating spatial cognition into LLM-based mobility agents.

  2. Extensive experimental validation. The authors conduct a comprehensive set of experiments on datasets from both New York City and Beijing. The reported results demonstrate consistent improvements across multiple evaluation metrics, providing encouraging evidence for the effectiveness of the proposed framework.

  3. Evaluation of practical downstream utility. Beyond conventional mobility simulation metrics, the authors investigate whether the simulated trajectories can reproduce meaningful real-world phenomena through segregation analysis. This additional evaluation is valuable, as it helps connect improvements in trajectory simulation with potentially meaningful urban and social patterns.

  4. A meaningful cognitive map design. The proposed three-layer cognitive map, together with its dynamic updating and intent-guided retrieval mechanisms, provides a reasonably intuitive and structured approach to modeling agents' spatial knowledge. I find this design practically motivated and potentially useful for constructing more realistic mobility agents.

Critical weaknesses
  1. Limited dataset diversity and insufficient evaluation on semantically rich mobility data. The experimental evaluation is restricted to two cities, Beijing and New York City, which limits the evidence for generalization across different urban environments and mobility patterns.

More importantly, I am curious why the authors do not consider semantically rich mobility datasets, such as Foursquare check-in data. Since the proposed framework explicitly models cognitive maps and intent-guided spatial knowledge retrieval, datasets containing semantic information about visited locations and points of interest would seem particularly relevant. Such evaluations could provide stronger evidence that the proposed cognitive map captures meaningful spatial knowledge rather than merely improving trajectory-level statistical consistency.

  1. Limited methodological novelty beyond prompt-based agent design. Although the conceptual motivation is interesting, the overall implementation appears to rely substantially on LLM-based agents, structured memory, and prompt-guided decision-making. These components are closely related to established approaches in LLM agent systems.

As a result, the work gives the impression of following an earlier generation of prompt-based agent frameworks rather than introducing a substantial methodological advance. While the cognitive map formulation itself is meaningful, I would like to see a clearer explanation of which components represent genuinely new technical contributions beyond conventional memory-augmented agents and retrieval-based prompting.

  1. Lack of comparison with established trajectory generation methods. The paper focuses on LLM-based mobility simulation but does not appear to include comparisons with representative trajectory generation models, such as DiffTraj [1].

Since the ultimate objective is to generate realistic human mobility trajectories, comparisons with established generative approaches would provide an important perspective on the actual quality of the simulated data. Without these comparisons, it is difficult to assess whether the proposed agent-based framework offers meaningful advantages over existing trajectory generation models in terms of trajectory realism, distributional fidelity, and downstream utility.

[1] Zhu, Y., Ye, Y., Zhang, S., Zhao, X., & Yu, J. J. Q. (2023). DiffTraj: Generating GPS Trajectory with Diffusion Probabilistic Model. Advances in Neural Information Processing Systems (NeurIPS 2023), 36, 65168–65188.

Questions for authors
  1. On dataset selection: Why do the authors evaluate CogMob only on the Beijing and NYC datasets? Have the authors considered additional semantically rich datasets, such as Foursquare check-in data, which could be particularly relevant for evaluating the proposed cognitive map and intent-guided retrieval mechanisms?

  2. On methodological novelty: Could the authors clarify the key technical differences between CogMob and existing memory-augmented LLM agent frameworks? In particular, which aspects of the three-layer cognitive map and its updating mechanism cannot be achieved through conventional structured memory and retrieval-augmented prompting?

  3. On comparisons with trajectory generation models: Could the authors explain why representative trajectory generation approaches, such as DiffTraj, are not included in the experimental comparisons? How does CogMob compare with these methods in terms of trajectory realism and distributional similarity when evaluated under comparable data conditions?

  4. On the contribution of cognitive map components: Could the authors provide further analysis of the individual contributions of the three cognitive map layers, the dynamic updating mechanism, and intent-guided retrieval? In particular, how much of the improvement comes from explicitly modeling spatial cognition, as opposed to simply providing agents with additional memory or contextual information?

7 comments
AuthorsOpenReview ↗

Response to Reviewer bocW (1/5)

Q1: On dataset selection: Why do the authors evaluate CogMob only on the Beijing and NYC datasets? Have the authors considered additional semantically rich datasets, such as Foursquare check-in data, which could be particularly relevant for evaluating the proposed cognitive map and intent-guided retrieval mechanisms?

Response: We thank the reviewer for this suggestion. We would like to clarify that the NYC dataset used in our experiments is already Foursquare check-in data, as stated in both the main text and the appendix. Thus, our evaluation already includes the type of semantically rich dataset suggested by the reviewer.

Semantic evaluation: To further evaluate the semantic fidelity of the generated mobility patterns, we follow the evaluation in [1,2,3,4] and use three metrics: Visit Purpose Distribution (VPD) measures the overall distribution of activity categories; Activity Transition Matrix (ATM) measures the conditional probability of the next activity category given the current one; and Daily Activity Routine Distribution (DARD) measures the joint distribution of activity categories and time of day. We map the POI categories in the NYC check-in data and the available activity labels in the simulated trajectories into six shared categories: Food, Shopping, Services, Health, Leisure and Sports, and Arts and Culture. DARD uses two-hour bins based on check-in times and simulated activity start times. All three metrics are compared using Jensen–Shannon divergence, with lower values indicating better performance. As shown in the table, CogMob achieves the best performance across all three metrics.

MethodVPDATMDARD
UML0.09230.18470.2568
LLMob0.07990.12500.2071
CoPB0.13130.28540.3037
CitySim0.10560.23590.2863
CogMob0.07290.10360.1607

Dataset diversity and cross-city generalization: We want to further emphasize that Beijing and NYC datasets are from two countries with significantly different geo characteristics, which have been proved useful and representative in many existing studies such as MIRAGE and CoPB. Nevertheless, to strengthen the generalization evidence, we have added an experiment on the Tokyo Foursquare POI check-in dataset [5]. Following the same procedure used for NYC, we generate agent profiles based on census information. We simulate the mobility of 1,000 agents throughout June 2012 and evaluate the generated trajectories against all real check-in records observed during the same month. Because generating trajectories with LLM-based baselines takes considerable time, we selected CoPB, the best-performing LLM baseline on the Beijing and NYC datasets, for comparison within the rebuttal period. We also included MIRAGE and TimeGeo, two non-LLM methods that performed well on both datasets and require training.

As shown in the table, CogMob outperforms TimeGeo, MIRAGE, and CoPB on seven, nine, and eight of the ten metrics in Tokyo, respectively, achieving average relative improvements of 8.66%, 16.75%, and 32.38% across all ten metrics. These results provide further support for CogMob’s generalizability across different urban environments.

DatasetMethodRadiusJumpDurationEntropyExploreReturnVisit-FreqCapacityLocFreqOdSim
TokyoTimeGeo0.09970.01420.23300.31100.23900.21570.26130.00290.03440.2425
MIRAGE0.19530.19440.23320.10230.05600.14890.14550.00360.02570.4079
CoPB0.02350.01990.33570.26020.20040.30980.62680.00900.07660.3476
CogMob0.02320.04030.22000.09930.05240.06270.03580.00020.09290.3310

[1] Santos, Gustavo H., Aline Carneiro Viana, and Thiago H. Silva. "When plausible is not realistic: Evaluating human mobility in llm-based urban simulation." arXiv preprint arXiv:2606.13835 (2026).

[2] Santos, Gustavo H., Aline Carneiro Viana, and Thiago H. Silva. "CityBehavEx: A Scalable and Empirically Validated LLM-Assisted Urban Simulation Platform." (2026).

[3] Wang, Jiawei, et al. "Large language models as urban residents: An llm agent framework for personal mobility generation." Advances in Neural Information Processing Systems 37 (2024): 124547-124574.

[4] Bougie, Nicolas, and Narimawa Watanabe. "Citysim: Modeling urban behaviors and city dynamics with large-scale llm-driven agent simulation." Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track. 2025.

[5] Yang, Dingqi, et al. "Modeling user activity preference by leveraging user spatial temporal characteristics in LBSNs." IEEE Transactions on Systems, Man, and Cybernetics: Systems 45.1 (2014): 129-142.

AuthorsOpenReview ↗

Response to Reviewer bocW (2/5)

Q2.A: On methodological novelty: Could the authors clarify the key technical differences between CogMob and existing memory-augmented LLM agent frameworks?

Response: We thank the reviewer for this question. We clarify the key technical differences between CogMob and existing memory-augmented LLM agent frameworks by considering two categories of prior work:

(1) General agent frameworks. These general-purpose agent frameworks [6-7] primarily store past events, interactions, knowledge, or skills in memory, without explicitly modeling spatial memory. Our ablations that remove the cognitive map while retaining episodic memory (CogMob w/o CM-A and CogMob w/o CM-B (Section 4.4 of the original paper)) show substantial performance declines relative to the full model.

(2) Agent frameworks for spatial reasoning. These frameworks include navigation agents [9-10] that build spatial maps or graphs and generative agents [8] that use environment trees to support activities and social interactions in virtual towns.

In terms of purpose, navigation frameworks build spatial knowledge to help agents reach task-specified goals. Park et al.’s environment tree helps agents choose where to carry out activities in a predefined town with only 19 top-level venues, with several categories represented by a single venue, such as a café, pub, or park. CogMob focuses on how individuals choose among the many destinations available for the same activity in a city, which places they repeatedly return to, and when they explore new destinations, based on their evolving spatial knowledge and experience.

Methodologically, navigation frameworks integrate observations into spatial maps or graphs and use the stored spatial relations to plan routes or select movements toward a task-specified goal. In Park et al.’s approach, agents add newly perceived venues, subareas, and objects to a partial environment tree with a predefined containment hierarchy. Overall, these approaches expand spatial knowledge through perception and use it to guide navigation or select where to carry out activities. However, simply accumulating observations alone does not fully capture how individuals understand and use urban space. Even people who visit the same places may form different impressions, remember different nearby POIs, perceive accessibility differently, and group places into different personal regions. These differences shape which destinations they consider, revisit, or explore.

In contrast, CogMob organizes spatial knowledge into places, routes, and regions. It distinguishes visited places from those the agent only knows about and forms personal regions by clustering visited places. This knowledge evolves through information exposure and experience. Agents learn about unvisited places through indirect information relevant to their interests and current intent. Reflection revises personal impressions of visited places, and memory strength decays over time. These updates change what the agent remembers, how it perceives places, and how its spatial knowledge is organized. CogMob then uses this evolving knowledge to support destination choice among urban POIs.

Experimentally, we introduce a variant that replaces the cognitive map with a flat place memory that expands through exploration and perception (i.e., CogMob with Perception-expanded Flat Memory (CogMob w PFM)). After each visit, it records the visited place and all nearby POIs, adapting the observation-based knowledge accumulation used in navigation maps and environment trees to the POI level. All other components and experimental settings remain unchanged. As shown in the table, CogMob achieves an average improvement of 33.36% over this variant on the Beijing dataset.

DatasetMethodRadiusJumpDurationEntropyExploreReturnVisit-FreqCapacityLocFreqOdSim
BeijingCogMob w PFM0.03250.03700.03730.14820.05430.09070.20280.00090.18780.3435
CogMob0.02140.02920.03780.13310.02710.02950.09110.00030.15440.2943
Improvement34.15%21.08%-1.34%10.19%50.07%67.47%55.07%64.79%17.78%14.32%

[6] Zhong et al. MemoryBank: Enhancing Large Language Models with Long-Term Memory. AAAI, 2024.

[7] Huang et al. A Survey of Agent Memory in the Second Half: Towards Self-Evolving and Long-Horizon Agents. TMLR, 2026.

[8] Park et al. Generative Agents: Interactive Simulacra of Human Behavior. UIST, 2023.

[9] Wang et al. GridMM: Grid Memory Map for Vision-and-Language Navigation. ICCV, 2023.

[10] Chen et al. MapGPT: Map-Guided Prompting with Adaptive Path Planning for Vision-and-Language Navigation. ACL, 2024.

AuthorsOpenReview ↗

Response to Reviewer bocW (3/5)

Q2.B: In particular, which aspects of the three-layer cognitive map and its updating mechanism cannot be achieved through conventional structured memory and retrieval-augmented prompting?

Response: We thank the reviewer for this question. Conventional structured memory and retrieval-augmented prompting can be implemented in different ways for mobility simulation. We consider two possible configurations:

  • Configuration 1: Episodic memory with external POI retrieval. The agent maintains episodic memory of its mobility events and retrieves candidate destinations from an external POI database when deciding where to go.
  • Configuration 2: Place memory with memory-based retrieval. The agent maintains a collection of visited and encountered places, expanding and updating this memory through visits, and perception. Candidate destinations are retrieved from this personal memory.

We first analyze the methodological limitations of the two configurations:

Limitation of configuration 1. Although their retrieval results may differ, candidate generation is not guided by a persistent personal spatial representation, which may limit its ability to reproduce observed mobility characteristics and dynamics.

Limitations of configuration 2. Flat place memory records locations an agent has visited or encountered, but a place inventory alone can’t fully capture its personal understanding of the city. Even people who have visited the same places may differ in their familiarity with those places, their subjective impressions, and which nearby destinations they readily recall. Moreover, individual place records do not explicitly represent experienced route connections or region-level spatial organization. These aspects shape which destinations an individual considers, revisits, or explores.

We then evaluate the two configurations by implementing each as a variant of CogMob:

  • CogMob w/o CM-B (Section 4.4 of the paper), corresponding to configuration 1. This variant retains episodic memory but removes the cognitive map, retrieving candidate destinations directly from the shared city-wide POI database. Candidate generation therefore lacks an explicit, persistent representation of the agent’s individual spatial knowledge and familiarity.
  • CogMob with Perception-expanded Flat Memory (CogMob w PFM), corresponding to configuration 2. This variant replaces the cognitive map with a flat place memory that expands through exploration and perception. After each visit, it records the visited place and all nearby POIs, retaining both visited locations and encountered but unvisited locations. Candidate destinations are retrieved from this memory, supplemented by external POI retrieval when insufficient candidates are available.

As shown in the table, CogMob improves eight of the ten metrics compared with w/o CM-B, with an average relative reduction in JSD of 30.12% across all ten metrics. Compared with the perception-expanded flat memory variant (CogMob w PFM), CogMob reduces JSD by an average of 33.36%.

DatasetMethodRadiusJumpDurationEntropyExploreReturnVisit-FreqCapacityLocFreqOdSim
BeijingCogMob w PFM0.03250.03700.03730.14820.05430.09070.20280.00090.18780.3435
CogMob0.02140.02920.03780.13310.02710.02950.09110.00030.15440.2943
Improvement34.15%21.08%-1.34%10.19%50.07%67.47%55.07%64.79%17.78%14.32%

Comments on the paper

3 comments

Dear Reviewers,

We have updated the manuscript to incorporate the revisions discussed in our rebuttal, with all changes highlighted in blue. We would appreciate it if you could review our responses and the revised manuscript. Please let us know if any concerns remain or if further clarification would be helpful. Thank you for your time and feedback.

Best regards,

The Authors

AuthorsOpenReview ↗

General Response to the Area Chair and All Reviewers (2/2)

During the rebuttal period, we addressed the following concerns raised by the reviewers.

Methodological novelty

  • Differences between CogMob and existing memory-augmented LLM agent frameworks. General-purpose agents store past experiences, while spatial agents build maps to navigate or select activity locations. For human mobility, however, individuals exposed to the same places may recall different destinations for the same activity. We clarified how CogMob captures this through evolving spatial knowledge beyond episodic memory and perception-based spatial representations. Two sets of ablations support the benefits of this design. CogMob w/o CM-A and CogMob w/o CM-B both perform worse than the full model despite retaining episodic memory. Compared with CogMob with Perception-expanded Flat Memory (CogMob w PFM), which records visited places and all nearby POIs, the full model reduces JSD by 33.36% on average across the ten Beijing metrics.

  • Differences between CogMob and existing personalized memory frameworks. We clarified how CogMob differs from GATSim, CitySim, and AgentMove by modeling evolving spatial knowledge of known but unvisited places, spatial connections, and regional organization. We also provided targeted ablation evidence showing average relative JSD reductions of 30.12% over the episodic-memory-only variant (CogMob w/o CM-B) in Beijing and 31.52% over the visited-only flat-memory variant (CogMob w VFM), with all other components unchanged in the latter comparison. These results demonstrate the benefits of cognitive maps beyond existing memory mechanisms.

Extended evaluation

  • Cross-city generalization. We extended our evaluation from Beijing and NYC to Tokyo, where we simulated 1,000 agents throughout June 2012 and compared their trajectories with Foursquare check-ins from the same month. CogMob achieves average relative reductions in JSD of 8.66% over TimeGeo, 16.75% over MIRAGE, and 32.38% over CoPB on the Tokyo dataset.
  • Cognitive-map ablations. We tested the effects of updates to route, region, and place knowledge, as well as forgetting and initialization. CogMob achieves average relative JSD reductions of 24.74%, 14.61%, and 20.30% over the corresponding ablation variants, respectively.
  • Semantic evaluation. We evaluated whether the simulated NYC trajectories capture which activities people undertake, how they transition between activities, and when those activities occur. CogMob achieved the lowest JSD among the evaluated methods on all three metrics: VPD, ATM, and DARD.
  • Additional training-based baseline. We adapted DiffTraj to stay-point data and evaluated it on Beijing and NYC. CogMob reduces JSD by an average of 60.47% on Beijing and 45.54% on NYC compared with DiffTraj.
  • Scalability. We reported average runtimes of 24.41 seconds per agent per simulated day on two A5000 GPUs and 7.55 seconds on one Pro 6000 GPU. When real-world mobility data are difficult to obtain, CogMob can generate high-quality synthetic mobility data at a manageable computational cost.

Clarifications

  • Segregation analysis. We clarified the experimental setup and the interpretation of the results.
  • Parameter choices. We explained how we selected the search radii, clustering radius, candidate limit, and memory decay rates.
  • NYC evaluation and variability across runs. We clarified how we evaluated simulated mobility against observed check-ins and how consistent the results were across runs.

We hope the Area Chair will consider these contributions and the substantial additional evidence provided during rebuttal when making the final recommendation. Thank you for your time and consideration.

Best regards,

The Authors

AuthorsOpenReview ↗

General Response to the Area Chair and All Reviewers (1/2)

Dear Area Chair and Reviewers,

We sincerely thank the reviewers for their constructive feedback. We have carefully addressed their concerns through substantial additional experiments, methodological clarifications, and revisions to the manuscript. We would like to summarize why we believe CogMob makes a meaningful contribution to ICLR.

Methodological contribution that introduces personalized, structured, and evolving cognitive maps on spatial knowledge to guide LLM agents’ long-term mobility decisions. Our central contribution is a framework that enables LLM-based agents to maintain personalized, partial, and evolving spatial knowledge and use it to guide long-term decisions. Unlike conventional agent memory that primarily records past experiences, CogMob explicitly represents what an individual knows about an environment, how this knowledge evolves, and how it shapes future decisions. This is realized through a three-layer cognitive map, experience-driven knowledge updating, and intent-guided retrieval. These mechanisms provide a general approach to modeling the relationship between an agent’s accumulated knowledge and its sequential decision-making.

Extensive experimental evidence supporting the proposed mechanisms. Compared with the strongest training-free LLM-based baselines, CogMob reduces JSD by an average of 49.8% in Beijing and 31.4% in NYC. In response to the reviewers, we added controlled ablations comparing CogMob with visited-only flat memory (CogMob w VFM) and perception-expanded flat memory (CogMob w PFM). CogMob achieves average relative JSD reductions of 31.52% and 33.36%, respectively. Additional component ablations support the contributions of dynamic updating, forgetting, and initial spatial knowledge. Furthermore, our plug-and-play experiment (Section 4.5) demonstrates that the proposed cognitive map can improve an existing LLM-based simulator, suggesting that its benefits extend beyond our specific implementation.

Enabling meaningful downstream scientific research. Importantly, our experiments demonstrate not only improved simulation accuracy but also the potential to enable new capabilities for downstream scientific research. We conduct a socioeconomic segregation study (Section 4.6 and Appendix C.4) to examine whether simulated mobility can reproduce social patterns documented in real-world empirical research. Our results reproduce both non-random socioeconomic segregation and the relationship between exploration behavior and experienced segregation, consistent with prior empirical findings. In contrast, the strong baseline CoPB struggles to reproduce these patterns, particularly the relationship between exploration and segregation (3D JSD: 0.1111 for CogMob vs. 0.6245 for CoPB). This distinction is especially important because the ability to reproduce meaningful social phenomena, rather than merely match conventional mobility statistics, is a key factor in determining whether researchers in transportation, social science, and urban planning can reliably use simulated data for scientific studies.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.