Aligning LLM agents with human learning and adjustment behavior: A dual agent approach

Personas, identity, and agents2026ElsevierApproved editorial review

Authors: Tianming Liu, Jirong Yang, Yafeng Yin, Manzi Li, Linghao Wang, Zheng Zhu

Keywords: LLM Traveler Agents, Learnable Persona, Online Calibration, Route Choice, Behavioral Simulation

Source: Open primary source (opens in a new tab)

6
Authors
6
Findings
8
Limitations
6
Evidence

Editorial summary

English

Published in Transportation Research Part C in 2026, this paper proposes calibrating GPT-4o traveler agents to predict how an individual changes routes in a repeated congestion game. Each traveler agent combines a four-period short memory, a 24-period long memory, and a free-form textual “persona” describing preferences, thresholds, and decision rules. A second GPT-4o agent acts as a calibrator. At each period it reevaluates the persona over an eight-observation window, writes textual “pseudo-gradients” attributing prediction errors, synthesizes update directions, generates candidate personas, re-simulates them, and retains a candidate only if it lowers loss. The result is then smoothed with a summary of up to 80 periods. Persona here is therefore a learned verbal policy for a binary task, not a validated psychological personality. The evaluation reuses a laboratory experiment with 15 students making 160 repeated choices between an expressway and one of two local roads. The first 80 periods are used for training and the last 80 for evaluation on the same individuals. The method reports 0.655 accuracy and 0.635 group-weighted F1, compared with 0.549 and 0.544 for the best baseline: absolute gains of 10.6 and 9.1 percentage points, or relative gains of 19.3% and 16.7%. It beats each baseline for 10 to 12 of the 15 participants depending on the metric, but is not best for everyone. On a two-dimensional behavior measure, switching after a loss and staying after a win, it reaches mean cosine similarity 0.974, only 0.004 above Bounded-LLM, and ranks last for three participants. This similarity is near saturation for every method, measures vector angle rather than magnitude, and does not observe cognition. In the aggregate simulation where all models receive the true previous system state, MAPE falls from 42.3% to 40.0% and the quantity called MSE from 7.49 to 5.94. However, the inspected arXiv formula averages an unsquared L2 norm, so it does not define mean squared error; missing code prevents resolving the discrepancy. The free-running simulation without true previous states is presented only through plots and qualitative discussion. An audit of the public input confirms 2,400 unique records, 15 people by 160 periods, with no missing participant-period pair. It also shows that Table 1 is mislabeled: its headings say first/last 20 days, but its percentages are calculated from the first/last 80 periods. Although the paper calls the input “real-world” data, it is a roughly 45-minute laboratory game. One of five available human groups is selected without a reported rationale. There are no significance tests, intervals, repeated GPT-4o runs, component ablations, or evaluation on unseen people, other models, or transport networks. The original human input is public, but the calibration code, prompts, exact GPT-4o snapshot, predictions, persona histories, losses, and cost accounting are not. The study supports improved temporal prediction for these 15 people in this setup, not recovery of their cognition or generalization to real travelers.

Español

Publicado en 2026 en Transportation Research Part C, este trabajo propone calibrar agentes viajeros basados en GPT-4o para que predigan cómo una persona cambia de ruta en un juego repetido de congestión. Cada agente viajero combina una memoria corta de cuatro periodos, una memoria larga de 24 y una «persona» textual libre que describe preferencias, umbrales y reglas de decisión. Un segundo agente GPT-4o actúa como calibrador: cada periodo reevalúa la persona sobre una ventana de ocho observaciones, redacta «pseudogradientes» textuales para explicar errores, sintetiza direcciones de cambio, genera personas candidatas, vuelve a simularlas y conserva una candidata solo si reduce el error. Después la suaviza con un resumen de hasta 80 periodos. La persona es, por tanto, una política verbal aprendida para una tarea binaria; no equivale a una personalidad psicológica validada. La evaluación reutiliza un experimento de laboratorio con 15 estudiantes que eligen durante 160 turnos entre una vía rápida y una de dos vías locales. Se usan los primeros 80 turnos como entrenamiento y los últimos 80 como evaluación para esas mismas personas. El método obtiene accuracy 0,655 y F1 ponderado por grupo 0,635, frente a 0,549 y 0,544 del mejor baseline: mejoras absolutas de 10,6 y 9,1 puntos, o relativas de 19,3 % y 16,7 %. Supera a cada baseline para entre 10 y 12 de los 15 participantes según la métrica, pero no es mejor para todos. En una medida bidimensional de comportamiento, cambiar tras perder y mantenerse tras ganar, alcanza similitud coseno media 0,974, apenas 0,004 sobre Bounded-LLM, y queda último para tres participantes. Esta similitud está casi saturada en todos los métodos, mide solo el ángulo del vector y no observa un mecanismo cognitivo. En la simulación agregada que proporciona a todos los modelos el estado real anterior, el MAPE baja de 42,3 % a 40,0 % y la cantidad denominada MSE de 7,49 a 5,94. Sin embargo, la fórmula de la versión arXiv inspeccionada promedia una norma L2 sin elevarla al cuadrado, por lo que no define un error cuadrático medio; el código ausente impide resolver la discrepancia. La simulación libre, sin estados reales previos, solo se presenta mediante gráficas y comentarios cualitativos. La auditoría del dato público confirma 2.400 registros únicos, 15 personas por 160 turnos y ningún hueco. También revela que la Tabla 1 está mal rotulada: dice «primeros/últimos 20 días», pero sus porcentajes se calculan con los primeros/últimos 80 turnos. Aunque el artículo habla de datos «reales», se trata de un juego de laboratorio de unos 45 minutos; de los cinco grupos humanos disponibles se selecciona uno sin justificarlo. No hay pruebas estadísticas, intervalos, repeticiones de GPT-4o, ablaciones ni evaluación con personas nuevas, otros modelos o redes de transporte. El dato humano original es público, pero no se han publicado el código de calibración, prompts, versión exacta de GPT-4o, predicciones, historiales de personas, pérdidas ni costes. El estudio respalda una mejora de predicción temporal para estas 15 personas y esta configuración, no la recuperación de su cognición ni la generalización a viajeros reales.

Research question

Can an LLM-based calibrating agent learn and update an individual textual persona so that another LLM agent better predicts the evolution of the route choices of human travelers?

Method

Fifteen GPT-4o traveler agents receive a textual persona and memories of 4 and 24 periods. Another GPT-4o calibrates each persona through moving window evaluation, textual error attributions, candidate generation and selection, and smoothing with long context. It is compared with Base-LLM, Recursive-LLM, and Bounded-LLM over an 80/80 temporal split of the same human group, and individual decisions, a fit vector, and aggregate flows are evaluated.

Sample: A selected group of 15 students, 9 in the OD with location 1 and 6 in the OD with location 2, makes 160 choices in a laboratory game. The analysis uses 1,200 observations from turns 1-80 for training and 1,200 from turns 81-160 for evaluation over the same individuals.

Findings

  • The method reaches 0.655 accuracy and 0.635 F1, compared to 0.549 and 0.544 for Base-LLM; these are absolute improvements of 10.6 and 9.1 points.
  • It outperforms each baseline for between 10/15 and 12/15 participants, but the comparisons lack uncertainty and it does not win for all individuals.
  • The mean cosine similarity of the behavior vector is 0.974 compared to 0.970 for the closest baseline; the method ranks last in three cases.
  • With controlled dynamics, MAPE goes from 42.3% to 40.0% and the quantity called MSE from 7.49 to 5.94; the published formula does not correspond to an MSE.
  • The free simulation qualitatively reproduces some trends, but has no quantitative table or complete baseline comparison.
  • The public data confirms the integrity of the group used and reveals that Table 1 summarizes 80+80 turns despite labeling them as 20+20 days.

Limitations

  • Only one of five available human groups is used, with 15 students, and the selection is not justified nor tested on the other groups.
  • The temporal split keeps the same individuals; it does not demonstrate generalization to new individuals, other environments, or natural mobility.
  • Only GPT-4o is evaluated and no snapshot, temperature, seeds, or repetitions are specified.
  • There are no significance tests, intervals, or a model of temporal and social dependence for the 1,200 evaluation decisions.
  • There are no ablations of persona, memories, pseudogradients, candidate selection, or smoothing.
  • The two-dimensional cosine similarity is close to the ceiling and does not validate cognitive mechanisms.
  • The MSE formula and the temporal label of Table 1 are inconsistent with their definitions and data.
  • Code, prompts, predictions, complete personas per period, raw results, and cost accounting for the new method are missing.

What the study does not establish

  • It does not demonstrate that the textual persona is a psychological personality or the real cognitive mechanism of the participant.
  • It does not test statistical significance despite the use of the term significant.
  • It does not demonstrate generalization to unobserved travelers, the other four groups, other LLMs, or real networks.
  • It does not allow interpreting the quantitative aggregate improvement as the result of a fully free simulation.
  • It does not certify that the value called MSE is a mean squared error under the published formula.
  • It does not offer an integral reproduction of the method with public artifacts.

Traceability

Scope: Full text

Version: arXiv:2511.00993v1; journal publication metadata checked for TRC 2026

Consulted source: https://arxiv.org/pdf/2511.00993

Review: Codex 32-page visual, current journal-publication, full-method, 15-participant x 160-period public-data, temporal-label, metric-formula, benchmark, statistical-design, construct-validity, artifact-search and claim-boundary audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4o as traveler-agent core
  • GPT-4o as calibration, critique, integration, editing and smoothing core

Instruments and metrics

  • Online textual pseudo-gradient persona optimization
  • Short-term and long-term memory retrieval
  • Binary choice accuracy
  • OD-group-weighted F1
  • Cosine similarity of switch-after-loss/stay-after-win vectors
  • Route-flow MAPE
  • Quantity labeled MSE with a formula inconsistency

Data used

  • Human group 1 from Comparing AI and human decision-making mechanisms in daily collaborative experiments
  • Zenodo 10.5281/zenodo.15307283
  • Controlled two-OD repeated route-choice laboratory game

Evidence and location

  • Traveler agent and persona architecture: arXiv v1 sections 3.1-3.2, pp. 6-10
  • Calibration by pseudogradients, candidates, and smoothing: arXiv v1 section 3.3 and Algorithm 2, pp. 10-13
  • Sample, 80/80 split, GPT-4o, and metrics: arXiv v1 sections 4.1-4.3 and Table 1, pp. 13-16
  • Individual results and learned personas: arXiv v1 Tables 2-6, pp. 17-24
  • Controlled aggregate results and free simulation: arXiv v1 Figures 5-6 and Table 7, pp. 24-28
  • Publication audit, public data, 20/80 label, MSE formula, artifacts, validity, and claim boundaries: reports/verification/article-258-trc-dual-agent-traveler-persona-data-metric-artifact-and-claim-audit.json