Skip to content

ELLMob: Event-Driven Human Mobility Generation with Self-Aligned LLM Framework

Conference: ICLR 2026 arXiv: 2603.07946 Code: GitHub Area: LLM/NLP Keywords: Human mobility generation, event-driven trajectory, LLM self-alignment, fuzzy-trace theory, cognitive decision-making

TL;DR

This paper proposes ELLMob, a framework grounded in Fuzzy-Trace Theory (FTT) from cognitive psychology. By extracting and iteratively aligning "habit gist" and "event gist," the framework reconciles the competition between users' routine patterns and social event constraints, enabling interpretable event-driven trajectory generation.

Background & Motivation

Human mobility generation aims to synthesize plausible spatiotemporal trajectory data, with broad applications in urban planning, traffic management, and public health. While LLMs have achieved success in routine trajectory generation, two critical challenges remain:

  1. Evaluation bias from data scarcity: Existing methods are predominantly developed and evaluated on non-event-day (stable-period) data, raising doubts about their reliability under sudden social events (natural disasters, public health emergencies).
  2. Lack of competing-decision reconciliation: Real-world mobility during events combines habitual regularity with shock-induced deviation—users retain routine activities at key anchor points (e.g., workplaces) while adjusting other behaviors. Existing methods either default to habitual patterns or are dominated by event constraints.

Concrete manifestations: - Typhoon: moving away from coastal areas, canceling non-essential commutes - COVID-19: self-restricting activity range - Olympics: restricted zones and traffic congestion

Method

Overall Architecture

ELLMob consists of three interconnected modules: (1) Event Schema Construction, which structures raw event narratives; (2) a trajectory generation module that leverages LLMs to produce candidate trajectories; and (3) gist-based reflective self-alignment, which iteratively reconciles competing decisions.

Key Designs

Event Schema Construction:

Free-text event descriptions are transformed into structured representations along four dimensions: - Event Profile: type, name, occurrence time, affected area - Intensity & Scale: quantitative indicators such as wind speed and precipitation - Infrastructure Impact: transportation and public facility operational status - Official Directives: government orders, applicable populations, and geographic scope

Three-Type Gist Extraction Based on Fuzzy-Trace Theory:

Gist Type Attribute Description Example
Pattern Gist Core behavior Primary behavioral pattern Daily commute to office
Inertial anchors Deeply embedded, non-negotiable components Returning home to a specific neighborhood at night
Vulnerability points Critical dependencies and single points of failure Reliance on a single railway line that may be suspended
Event Gist Primary intent Core impact of the event on mobility decisions High outdoor risk; strong incentive to stay home
Behavioral influence Survival, social dynamics, and compliance Evacuating from coastal areas, seeking indoor shelter
Risk-benefit assessment Cost-benefit analysis of event-related risks Injury risk outweighs benefit of non-essential outings
Action Gist Primary intent Main purpose driving trajectory choices Procuring necessities from a nearby store
Habit adherence Degree to which habitual patterns are retained Low: deviating from usual work commute
Event compliance Degree to which event constraints are followed High: short trips avoiding hazardous areas

Reflection-based Alignment:

A two-stage iterative process:

  1. Alignment Auditing: Candidate trajectories are examined along two binary dimensions:

    • Internal Alignment: Does the trajectory reflect the user's intrinsic habitual mobility patterns?
    • External Alignment: Does the trajectory represent a reasonable, compliant response to event constraints?
    • A trajectory is accepted only when both criteria are satisfied.

  2. Corrective Refinement: Upon failure, precise failure reasons are provided as feedback to guide regeneration. A maximum of \(K=3\) iterations is allowed; upon timeout, the most recent valid trajectory from a buffer is used and unmet constraints are reported.

Loss & Training

  • Primary backbone: GPT-4o-mini (2025-01-01-preview)
  • Temperature 0.1, Top-p 1
  • Trajectory modeling at 10-minute temporal resolution
  • Spatial grid parameter \(S = 10\)
  • Maximum iterations \(K = 3\) (determined via parameter study)

Problem Formulation:

\[F: (D_{\text{long-term}}^{(u)}, D_{\text{short-term}}^{(u)}, E_{ctx}) \mapsto \tau\]
  • Long-term trajectory \(D_{\text{long-term}}^{(u)}\): historical trajectories from an earlier pre-event period
  • Short-term trajectory \(D_{\text{short-term}}^{(u)}\): recent pre-event trajectories
  • Event context \(E_{ctx}\): structured event schema

Key Experimental Results

Main Results

Method comparison across three events (JSD↓, lower is better):

Model Typhoon SI Typhoon SD Typhoon CD Typhoon SGD
LSTM 0.1336 0.1039 0.0555 0.1111
DeepMove 0.1697 0.0826 0.0266 0.0759
LLM-MOB 0.1214 0.0468 0.0285 0.0344
LLM-Move 0.1267 0.0392 0.0136 0.0303
LLMOB 0.0949 0.1195 0.0123 0.0256
ELLMob 0.0642 0.0200 0.0041 0.0173
Model COVID SI COVID SD COVID CD COVID SGD
LLM-MOB 0.1166 0.0532 0.0234 0.0353
LLM-Move 0.1408 0.0567 0.0127 0.0503
LLMOB 0.1013 0.1051 0.0186 0.0286
ELLMob 0.1003 0.0444 0.0080 0.0268
Model Olympics SI Olympics SD Olympics CD Olympics SGD
LLMOB 0.0973 0.0274 0.0110 0.0051
LLM-Move 0.1967 0.0298 0.0101 0.0057
ELLMob 0.0617 0.0061 0.0022 0.0035

Key figures: ELLMob outperforms the strongest baseline by 32.3% on SI under the typhoon scenario and by 16.5% on SD under COVID-19, with an average improvement of 46.9% over the strongest baseline.

Ablation Study

Variant Typhoon SI Typhoon SD COVID SI COVID SD
Full ELLMob 0.0642 0.0200 0.1003 0.0444
w/o I.A.&E.A. 0.1304 0.1270 0.2331 0.1077
w/o I.A. (E.A. only) 0.0835 0.0720 0.1235 0.0950
w/o E.A. (I.A. only) 0.0680 0.0258 0.2237 0.0860
w/o Eve. Ext. 0.0736 0.0273 0.2037 0.0741

Key ablation findings: - Removing external alignment causes a 132.4% degradation on SI in the COVID-19 scenario—external alignment is critical for handling significant behavioral deviations. - Removing internal alignment causes the model to over-correct (e.g., unreasonably increasing health-care-related trips). - Cognitive self-alignment improves non-aligned variants by an average of 69.5%.

Key Findings

  1. LLM-based methods consistently outperform deep learning methods, particularly on spatial consistency metrics (SD, SGD), owing to their ability to integrate event context.
  2. Existing LLM baselines fail severely in event scenarios: they either default to habitual patterns (underestimating health-related mobility) or over-respond to event constraints (completely suppressing social activity).
  3. Disaster activity classification: ELLMob achieves the highest F1-score in identifying active users during typhoons (binary classification), with a recall of 59.3%.
  4. Internal and external alignment serve distinct roles: internal alignment provides foundational plausibility, while external alignment delivers scenario-specific correction.

Highlights & Insights

  1. Cognitively-grounded AI framework design: Fuzzy-Trace Theory (FTT) is incorporated into LLM-based trajectory generation—this represents principled architectural design with a cognitive science foundation rather than simple prompt engineering.
  2. First event-annotated mobility dataset: Covering three distinct event types (natural disaster / public health emergency / major sporting event), filling a significant data gap.
  3. Explicit reconciliation of competing decisions: Trajectory generation is reframed from "maximizing statistical likelihood" to "cognitive plausibility," making the decision process traceable through gist alignment.
  4. Comprehensive experimental coverage: 12 baseline methods (6 deep learning + 4 LLM + ablation variants), 4 evaluation metrics, and 4 scenarios.
  5. The average improvement of 46.9% is substantial, maintained across all three event types.

Limitations & Future Work

  1. Geographic limitation of data: Only Twitter/Foursquare check-in data from the Greater Tokyo Area is used; generalizability remains to be verified (though supplementary experiments in Osaka are provided in the appendix).
  2. LLM API cost: The iterative alignment process requires multiple API calls, resulting in high inference overhead.
  3. Manual design of event schema: The four-dimensional event schema relies on domain expertise, limiting automation.
  4. Sparsity and bias of check-in data: Social media check-ins cannot fully reflect real-world mobility.
  5. Coarse temporal resolution: A 10-minute resolution may fail to capture fine-grained behavioral changes.
  • LLM-MOB (Wang et al., 2023), LLM-Move (Feng et al., 2024), and LLMOB (Wang et al., 2024) serve as the primary LLM baselines.
  • Fuzzy-Trace Theory (Reyna & Brainerd, 1995) provides the cognitive theoretical foundation—the fact that gist is expressible in language makes the integration of FTT with LLMs feasible.
  • Self-alignment/self-reflection: Unlike general self-alignment methods for hallucination correction, the self-alignment in this paper focuses specifically on reconciling competing decisions.
  • Insight: Cognitive science theories can provide principled guidance for LLM application architecture design, rather than relying solely on large-scale prompt engineering.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ — First event-driven mobility generation framework; FTT-gist alignment is a distinctively original design concept.
  • Technical Depth: ⭐⭐⭐⭐ — The integration of cognitive theory with LLMs is rigorous, and the problem formalization is clear.
  • Experimental Thoroughness: ⭐⭐⭐⭐ — 12 baselines, 4 scenarios, multi-dimensional evaluation, and comprehensive ablation.
  • Practicality: ⭐⭐⭐⭐ — Direct application value for emergency management and urban planning.
  • Writing Quality: ⭐⭐⭐⭐ — Framework diagrams are clear; the introduction of cognitive theory is well-executed.

Overall: ⭐⭐⭐⭐ (4.5/5) — A highly creative interdisciplinary work that organically integrates cognitive psychology with LLM-based trajectory generation. The problem definition is novel, experimental performance is outstanding, and the paper stands as an excellent representative of the LLM-for-Science research direction.