PaperlayerView exact revision on arXiv

Source-linked AI summary

An LLM-powered Agent Framework for Heterogeneous Evacuation Behavior Modeling under a Moving Threat in a Public Plaza

Jian Ma, Runxin Yu, Tianyu Tang, Xiaolian Li

arXiv:2609.37009v1physics.soc-phcs.MA

TL;DR

Evacuation under a moving threat requires modeling how people interpret incomplete information, remember experience, and assess evidence. This paper develops an LLM-powered agent framework linking those processes to choices and physical movement. Usable-exit knowledge was strongly associated with evacuation, while evacuation time was associated with geometry, information, and affect together.

  • Problem

    Modeling evacuation under a moving threat requires representing how individuals interpret changing information, remember experience, assess danger, and choose routes.

    Source: §1 · pp. 1–2

  • Method

    The framework links private perception, personality, retained experience, and decisions to physically executed movement, with memory revisions connected to choices.

    Source: §1 · p. 2, §6 · p. 18

  • Results

    89.5% of pedestrians with knowledge of a usable exit were rescued, compared with 1.05% without such knowledge; evacuation time was associated with geometry, information, and affect together.

    Source: §6 · pp. 18–19

  • Takeaways & Limitations

    The framework makes the formation, continuation, and revision of escape plans accessible to analysis through linked memory revisions, decisions, and movement.

    Source: §6 · p. 18

    Whether model-generated interpretations correspond to human judgments, route choices, and responses to communicated information remains to be established.

    Source: §5.3 · p. 17

Abstract

from arXiv · show

Modeling heterogeneous evacuation behavior under a moving threat is difficult because human perception, memory, and evidence evaluation are not well captured by fixed rules. We propose a novel LLM-powered agent-based framework to represent these internal decision processes. Each pedestrian agent perceives a private symbolic ASCII view, maintains a Memory-based Knowledge Graph derived solely from individual observations, and makes decisions through persona-conditioned prompts under a common sampling configuration. A compressed decision context with stateless memory preserves trial-and-error experience across turns while excluding reasoning traces, and a validation engine separates behavioral choice from physical feasibility by executing routes only over observed terrain. We evaluated eight personality compositions in eight paired randomized blocks within a simulated public plaza. Usable-exit knowledge was strongly associated with evacuation success: 89.5% of agents possessing such knowledge evacuated, compared with 1.05% of those without it. Personality compositions also differed in their evaluation of remembered threat evidence: the proportion of danger assessments varied by 0.265, while high-urgency, low-directness decisions ranged from 11.41% to 34.33%. After direct threat sightings, responses converged, with 99.6% of assessments classifying the situation as dangerous. Movement was selected in 99.8% of decisions. Overall, evacuation outcomes were strongly associated with information access, while evacuation time was jointly associated with spatial geometry, information, and affect. The framework provides an auditable approach to generating endogenous behavioral heterogeneity through persona-conditioned LLM agents in crowd-evacuation simulations.

1. Introduction

Evacuation under a moving threat requires people to interpret incomplete information, assess risk, and choose routes in spaces shaped by geometry. The paper develops an LLM-powered agent framework to examine how individual experience enters these decisions.

1. Introduction

  • A moving threat creates an evacuation problem in which information arrives unevenly and route choice depends on spatial cognition, risk perception, and scene geometry.People must determine what is happening, whether they are at risk, where exits are, and how nearby people may respond.Source: §1 · p. 1
  • Macroscopic and mesoscopic models represent crowds as flows or distributions over individual states, while microscopic models resolve individuals but generally leave cognition and perceived information outside the model.At these scales, heterogeneity is difficult to capture or is represented through a limited set of activity variables or state classes.Source: §1 · p. 1
  • Recent pedestrian and agent-based models make information and behavioral heterogeneity more explicit, but decisions remain encoded in objectives, learned functions, strategy sets, weighted functions, or prescribed rules.These approaches include changing access conditions, reduced visibility, dynamic hazards, and heterogeneous exit-choice strategies.Source: §1 · p. 2
  • The study asks how changing information is interpreted, retained, and translated into evacuation decisions as evidence about danger and available exits develops.It builds on LLM-agent work in evacuation and wayfinding and connects changing understanding to route choice and movement.Source: §1 · p. 2
  • The paper identifies four framework requirements: interpretable environmental input, endogenous population heterogeneity, persistent inspectable memory, and person-to-person information transfer.Prior approaches represent heterogeneity through mobility attributes, sensitivity parameters, decision mechanisms, or personality traits.Source: §1 · p. 2
  • The framework connects private observations, personality profiles, and retained experience to successive evacuation choices and physically executed movement.It reconstructs memory changes alongside choices, supporting analysis of how pedestrians interpret danger, maintain escape goals, and reconsider routes.Source: §1 · p. 2

2. Methodology

The framework separates each pedestrian’s perception, decision, memory updating, and physical execution. It bases route validation and movement on the pedestrian’s observed, passable terrain.

2. Methodology

  • Each pedestrian maintains a private known map containing observed terrain, and the engine validates routes through observed, passable cells.After a run, memory revisions and decision records are reconstructed as graphs for analysis.Source: §2 · pp. 2–3
  • Figure 1 presents the framework as distinct stages of local observation, decision, memory, and physical execution.Local observations update each pedestrian’s private known map.Source: Figure 1 · p. 3
  • The decision model receives current perception, persona, engine-computed affect, and short-term context, then selects an action and a route or perceived target.The engine generates and validates routes using the pedestrian’s known map before applying physical movement rules.Source: §2 · p. 3

2.1. Personality Modeling and Prompt Specialization

Personality enters the framework as a stable prompt-based profile that conditions, but does not prescribe, decisions. The decision process also incorporates observations, remembered experience, affect, and communicated information.

2.1.1. Personality Configuration

  • The framework uses five-factor personality scores—openness, conscientiousness, extraversion, agreeableness, and neuroticism—as part of each agent’s persistent persona description.T-scores map to five interpretive bands and are paired with salient facets and behavioral anchors; neuroticism also determines sensitivity and recovery coefficients.Source: §2.1.1 · pp. 3–4, §2.1 · p. 3

2.1.2. Behavioral Differentiation through Persona-Conditioned Decisions

  • A stable persona profile conditions how the language model interprets perception, affect, and remembered experience, without prescribing a fixed action.Agents presented with the same environmental evidence may therefore make different choices.Source: §2.1.2 · p. 4
  • As trajectories diverge, agents accumulate different observations and memory texts that can accompany differences in later decisions.The paper describes reconstructed graphs as records of this evolving history.Source: §2.1.2 · p. 4
  • The shared policy contains seven instruction blocks, while the identity section varies by agent; prompt context also includes updated perception, affect, and retained memory.Because neuroticism also affects engine-maintained fear, the design does not isolate the causal effect of persona text.Source: §2.1.2 · p. 4, Table 2 · p. 4
  • Each agent’s identity section combines demographic type and FFM profile with scores, bands, salient facets, and behavioral tendencies considered alongside observations, memory, and affect.Figure 2 depicts trait-score encoding while general decision instructions and sampling settings remain fixed across agents.Source: §2.1.2 · p. 4, Figure 2 · p. 5
  • The decision process links glyph interpretation and situation appraisal to action selection, allowing new evidence to prompt reconsideration of goals, routes, or targets.Directly seeing the threat or an injured person is specified as strong evidence of danger; the model supplies the appraisal label.Source: §2.1.2 · pp. 4–5
  • Anonymous shouts convey direction and approximate distance; listeners interpret them alongside observations, memory, personality, and affect before choosing a response.The engine records an updated appraisal as normal, unsure, or danger.Source: §2.1.2 · p. 5

2.2. ASCII-Based Local Perception and Spatial Representation

The section defines how agents receive a private symbolic view of the plaza, built from layered environment data and individual observations. It connects that view to memory, model-selected actions, and validation against observed terrain.

2.2.1. Layered Environment Representation

  • The private belief layer stores only personally observed terrain, while the effective layer governs physical collision and availability.Target selection and local route generation use the belief layer; physical resolution uses the effective layer.Source: §2.2.1 · p. 5

2.2.2. Field of View and the perception map

  • The perception map combines radius, view-cone, and line-of-sight visibility with eight-cell Moore awareness; initialization also reveals terrain within Chebyshev radius 3.Observed cells are recorded with their observation time in private memory, which supplies route options and supports target validation.Source: §2.2.2 · pp. 5–6, Figure 3 · p. 6

2.2.3. Symbolic Encoding of the Perceived Field

  • Unknown cells appear as ?, and the model returns an action, assessment, standing intent, and route identifier or target offset for validation.A* rejects unknown, blocked, or disconnected targets, and only validated belief-grid cells are executed.Source: §2.2.3 · p. 6

2.2.4. Properties of the Two-Dimensional Symbolic Field

  • Character grids preserve spatial adjacency under rotation, while a supplementary frozen-snapshot experiment compared four representations and kept equal-width ASCII fixed in the main experiment.The four supplementary representations were no map, structured cells, lean ASCII, and ASCII with a decoding protocol.Source: §2.2.4 · pp. 6–7

2.3. LLM-powered Agent Modeling Framework

The framework models pedestrian decisions from partial observations, persona, affect, and retained experience, while keeping behavioral choices distinct from physically validated movement. Separate decision and memory models support turn-by-turn choices and auditable changes in remembered facts.

2.3.1. Perception–Decision–Action Loop

  • The decision loop updates observations and applies validated choices at turn boundaries, while agents continue along existing legal routes during pending computation.A newly observed armed person or motionless body can replace an unfinished assessment; the adversary and movement are resolved after decisions are applied.Source: §2.3.1 · p. 7

2.3.2. Dual-Model Architecture and Decision Layers

  • Two models separate decision-making from memory revision, while three decision layers cover standing goals, route selection, and bounded route execution.The engine offers routes over the agent’s private knowledge graph and annotates observable route attributes; the model may choose a target if none is suitable.Source: §2.3.2 · p. 7

2.3.3. Context Compression and Experience Retention

  • Compressed operational context retains intentions, validated strategies, corrections, and unresolved questions, but excludes reasoning traces and stale details.Long-term memory is rewritten from current notes and the committed turn; omitted facts count as forgotten, and stale threat evidence retains its last observed cell and time.Source: §2.3.3 · p. 8

2.3.4. Graph-Structured Memory

  • The stateless memory model rewrites each agent’s record independently, and post-run graph reconstruction links remembered facts across revisions and to recorded decisions.Character-bigram Jaccard overlap matches reworded facts; a 0.40 threshold separates same-fact scores of 0.43–0.60 from distinct-fact scores of 0.00–0.11.Source: §2.3.4 · p. 8

2.3.5. Affective State

  • Engine-computed fear and model-returned appraisal are distinct affect channels; the appraisal reports situation and urgency, and may vary by language model.Fear responds to active stimuli and personality, while the decision prompt receives fear and its active stimuli; Figure 4 depicts versioned memory facts and recorded cross-track associations.Source: §2.3.5 · pp. 8–9, Figure 4 · p. 9

2.3.6. Time Step, Movement, and Contention

  • Agents move at most one grid cell per physical step, and contention resolution prevents overlap and other conflicting moves without changing stated preferences.Active pedestrians at the time limit are recorded as unresolved rather than as reached by the adversary.Source: §2.3.6 · pp. 9–10

2.3.7. Algorithmic Adversary and Evidence Propagation

  • A fixed, perception-limited patrol–pursuit policy governs the adversary, while threat evidence reaches pedestrians through local observation and communication.The adversary pursues the nearest visible pedestrian, searches toward the last observed cell after losing sight, and otherwise resumes patrol; shouts can reach listeners within 15 m without line of sight.Source: §2.3.7 · p. 10

3. Experimental Design

The experiment varies group personality composition while holding the plaza, movement, threat, communication, memory, and language-model conditions constant. Eight compositions are evaluated in eight matched blocks, using run-level outcomes and pedestrian-level records.

3.1.1. Spatial Layout and Emergency Opening Control

  • The plaza uses two entrance gates that close during the emergency and three eastern evacuation exits that remain open in the reference scenario.Eleven pedestrians occupy a 155 × 164 raster containing obstacles and a moving threat; openings are learned through direct observation or other pedestrians’ information.Source: §3 · p. 10, §3.1.1 · pp. 10–11

3. Experimental Design

  • Eight personality compositions were evaluated in eight matched blocks, with personality composition as the experimental factor and other simulation conditions held constant.The analysis covers situation appraisal, urgency and affect, objectives and routes, exit information, and evacuation outcomes.Source: §3 · p. 10

3.1.2. Outcome Measures

  • Outcome measures cover evacuation outcomes, threat exposure, and pedestrian behavior; movement directness is measured over the 10 seconds following each applied decision.The directness analysis includes 10,947 decisions from 704 pedestrians.Source: §3.1.2 · p. 11, §3.1.3 · p. 11, Table 4 · p. 11

3.1.3. Statistical Treatment

  • Homogeneous compositions are compared within paired blocks, while regression models summarize associations with evacuation status and log evacuation time.The regression factors include geometry, personality, information, and affect, fitted separately and together; the models describe associations within the recorded simulations.Source: §3.1.3 · p. 11

3.1.4. Language Model Configuration

  • Both decision and memory roles use GLM-5.3-flash at temperature 0.2, with higher reasoning for decisions triggered by direct visual evidence.Figure 5 depicts the plaza layout and three illustrative trajectories from one completed run.Source: §3.1.4 · p. 11, Figure 5 · p. 12

3.2. Experimental Conditions

  • Five homogeneous groups and three mixed groups define the eight compositions, including a rule-following to anxious-conformist dose ladder.The five homogeneous conditions are compared pairwise within blocks; a post hoc analysis measures movement relative to the last observed and actual threat positions.Source: §3.2 · p. 12

4. Results

Across 704 pedestrians, 462 (65.6%) evacuated and 202 (28.7%) were killed; analyses trace decision revision, personality-linked appraisal, memory, exit information, and evacuation time.

4.1.1. Decision Revisions during Evacuation

  • Across 10,947 decisions, plan-level revisions occurred in 64.5% of epochs, versus 8.8% goal revisions and 9.9% execution adjustments.Goal revisions accompanied absolute urgency changes of at least 0.15 in 58.6% of epochs, versus 4.3% for smaller changes; physical constraints accompanied revisions at plan level or above in 54.3% of decisions, while threat avoidance and movement toward a known exit persisted with probabilities of 0.982 and 0.931.Source: §4.1.1 · pp. 12–13

4.1.2. Personality, Appraisal, and Arousal

  • With remembered threat information, danger-assessment rates differed by 26.5 percentage points across personality profiles, versus at most 0.2 points without threat information.Direct sightings narrowed the gap to 3.3 percentage points across the five compositions.Source: §4.1.2 · p. 13
  • 31.6% of decisions among pedestrians who were killed combined high urgency with low directness, versus 7.0% among those who evacuated; shares across compositions ranged from 11.41% to 34.33%.The outcome comparison had an AUC of 0.839; calm-independent agents had the lowest share and rule-following agents the highest across the five homogeneous compositions.Source: §4.1.2 · p. 13
  • Across 102,045 memory operations, writing accounted for 33.8%, forgetting 29.9%, rewriting 24.7%, and explicit retention 11.6%.Anxious-conformist agents had a 15.6% explicit-retention share and 33.5% of decisions without revision; case records also show memories and plans changing after a threat sighting or recognition of a closed gate.Source: §4.1.2 · p. 14

4.1.3. Information Access and Survival

  • 89.5% of pedestrians with knowledge of a usable exit reached safety, compared with 1.05% of those without such knowledge.Rescue rates were 1.05% with no usable exit known, 64.3% with one, and 95.7% with two; none of 58 pedestrians whose only exit information concerned a closed gate was rescued.Source: §4.1.3 · p. 14
  • Shouts primarily carried alarm while usable-exit messages conveyed routes; exit information traveled through few senders and at most one hop, with 591 observing exits directly and 64 learning indirectly.Among 311 pedestrians whose previous appraisal was normal and whose first evaluated decision containing a shout had no direct or remembered threat evidence, 286 (92.0%) reported higher urgency than in their preceding decision; 254 (81.7%) assessed the situation as unsure.Source: §4.1.3 · p. 14

4.1.4. Factors Associated with Evacuation Time

  • The combined geometry, information, and affect model for evacuation time reached R^2 = 0.8738, compared with 0.0031 for personality composition.Separate model R^2 values were 0.6608 for geometry, 0.6864 for information, and 0.6859 for affect.Source: §4.1.4 · p. 15

5. Discussion

The discussion frames the framework as a way to connect individual interpretation and memory to movement. It also identifies the study’s scope and directions for testing these behavioral sequences and connecting them to crowd-level patterns.

5.1. Contribution to Evacuation Modeling

  • The framework makes agents’ situation descriptions, retained experience, and route revisions jointly inspectable alongside movement, including how different intentions can accompany similar trajectories.A continuing goal may be pursued through changing routes.Source: §5.1 · pp. 15–16
  • Across 10,947 decisions, individual high-urgency, low-directness shares distinguished death from evacuation (AUC 0.839); Figure 7 compares pooled urgency and directness and five compositions.The figure reports pooled medians of urgency 0.60 and directness 0.606.Source: Figure 7 · p. 16

5.2. Memory, Interpretation, and Route Revision

  • Memory cases show earlier cues can be reinterpreted and new exit knowledge assessed against an existing plan; the reconstructed graph links retained information to decisions.The recorded sequence supports analysis of how interpretations persist and plans are revised.Source: §5.2 · pp. 16–17

5.3. Scope and Limitations

  • The findings concern one small pedestrian population, plaza, and moving-threat policy; personality differences concern combined configurations, while correspondence with human judgments remains unestablished.The discussion also identifies human route choices and responses to communicated information as comparisons still to be established.Source: §5.3 · p. 17

5.4. Implications for Future Evacuation Research

  • Future studies could compare simulated responses with human behavior, test warning content, and connect individual decision histories to collective patterns such as aggregate exit demand.Proposed tests include when participants recognize danger, interpret uncertain exits, and abandon preferred routes.Source: §5.4 · pp. 17–18

6. Conclusions

The conclusions link evacuation outcomes to exit information and evacuation duration to geometry, information, and affect. They present the framework as a way to examine how individual plans develop and contribute to movement.

6. Conclusions

  • Linking memory revisions to decisions and movement makes the formation, continuation, and revision of escape plans accessible to analysis.The conclusions describe memory cases as hypotheses about delayed responses and persistence toward a known exit.Source: §6 · pp. 18–19
  • Most decision changes were plan-level, with physical constraints usually prompting goal-preserving adjustments; high-urgency, low-directness shares ranged from 11.41% to 34.33% across compositions.Large urgency changes and new direct threat evidence were associated with goal revisions.Source: §6 · p. 18
  • 89.5% of pedestrians with usable-exit knowledge were rescued, compared with 1.05% without it; evacuation time was associated with geometry, information, and affect together.The authors report a strong association between usable-exit knowledge and evacuation outcomes.Source: §6 · pp. 18–19

Supplementary Figures

Figure S1 depicts shout and exit-information transmission among pedestrians in one run and reports experiment-wide associations with outcomes.

Supplementary Figures

  • Figure S1 holds the same eleven pedestrians in place across panels and encodes them by outcome, with edges representing deduplicated sender–receiver pairs.The two panels show heard shouts and transmitted exit information, respectively.Source: Figure S1 · p. 20
  • The figure separates the run’s directed shout network from its directed exit-information network.Edges represent deduplicated sender–receiver pairs in each panel.Source: Figure S1 · p. 20
  • Across the experiment, shout out-degree had an AUC of 0.808 for being killed, while exit knowledge had an AUC of 0.651 for rescue.Source: Figure S1 · p. 20
Loading 2609.37009v1…