Towards Real-world Human Behavior Simulation: Benchmarking Large Language Models on Long-horizon, Cross-scenario, Heterogeneous Behavior Traces

Evaluation and psychometric validity2026arXivApproved editorial review

Authors: Jiawei Chen, Ruoxi Xu, Boxi Cao, Ruotong Pan, Yunfei Zhang, Yifei Hu, Yong Du, Tingting Gao, Yaojie Lu, Yingfei Sun, Xianpei Han, Le Sun, Xiangyu Wu, Hongyu Lin

Keywords: Persona conditioning, Human simulation, Safety and bias

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

14
Authors
9
Findings
10
Limitations
8
Evidence

Editorial summary

English

OmniBehavior proposes a behavior-simulation benchmark built from real Kuaishou traces. The paper describes 200 users, five scenarios, 22 action types, and 2.12 million interactions from September through November 2025. Each profile combines demographic and household attributes with histories that may include exact timestamps, video text, OCR, ASR, searches, products, advertisements, and customer-service conversations. Users are selected with K-means by taking the person nearest each of 200 centroids: this broadens diversity coverage but does not produce a population-representative sample or support prevalence estimates. The main task supplies a profile, history, and current context to predict binary actions, video duration, or the next customer-service reply. The primary setup uses September for history and October-November for testing, with a claimed 30 test actions per user and 6,000 cases. Eleven models are evaluated. Claude Opus 4.5 scores 44.55 on the published ranking, followed by GLM-4.7 at 41.46 and Claude Sonnet 4.5 at 40.49; Qwen3-235B scores 32.11. The overall score, however, is an unweighted mean of six heterogeneous components: binary-scene F1, 100 minus NMAE for duration, and a Claude-Sonnet-4.5 text score multiplied by ten. No intervals or tests accompany rank differences, and Claude-Sonnet-4.5 is simultaneously a competitor, text judge, and sentiment/style annotator without human validation. For memory, Qwen3-235B scores 24.27 with summarization, 21.13 with truncation, and 20.38 with RAG; these are averages across unlike scenes and do not support a general conclusion about all memory systems. Across 66 histories longer than 128k tokens, increasing context does not improve performance consistently. The paper calls 180 pre-conversion trajectories inferred by Claude and manually reviewed causal chains; because they come from observational logs without an intervention or counterfactual, they are retrospective attributions rather than causal effects. It also compares daily Jaccard drift in OmniBehavior, .6311, with LoCoMo, .1698, but platform, tasks, and generation differ and do not isolate a general real-versus-synthetic property. Its hyperactivity, utopian-bias, and personality-homogenization labels depend on balanced sampling, Claude annotations, and action-rate vectors; the last measures behavioral similarity, not validated personality. Artifact inconsistencies further constrain interpretation. Dividing 2.12 million interactions by 200 users gives 10,600 per user, whereas the reported 8,143 average implies 1,628,600, an unexplained difference of 491,400. In a deterministic audit of 20 pairs from the public snapshot, English JSON files contain 1,324 actions and Chinese files 2,688; only five pairs match, with the difference concentrated in video events, 355 versus 1,719. This establishes release drift in the audited sample, not a corpus-wide mismatch rate. The code requires at least 60 actions to construct 30 cases, while audited released users fall below that threshold, so the current artifact does not guarantee 200 users or 6,000 tasks; it also lacks reproducible configuration for all eleven models, outputs, judgments, and results. The dataset is public and ungated. Although direct identifiers are replaced, precise timestamps, demographics, and rare content remain longitudinal quasi-identifiers; consent, privacy evaluation, withdrawal, takedown contact, and anonymization performance are undocumented. The work contributes a rich industrial resource and inspectable code, but its findings should be read as results on a curated benchmark and an arXiv v2 preprint, not as representative simulation, causal evidence, or validated human personality measurement.

Español

OmniBehavior propone un benchmark de simulación de conducta construido con trazas reales de Kuaishou. El artículo describe 200 usuarios, cinco escenarios, 22 tipos de acción y 2,12 millones de interacciones entre septiembre y noviembre de 2025. Cada perfil combina datos demográficos y familiares con historiales que pueden incluir marcas temporales exactas, textos de vídeo, OCR, ASR, búsquedas, productos, anuncios y conversaciones de atención al cliente. Los usuarios se eligen mediante K-means y se toma el más cercano a cada uno de 200 centroides: esta decisión amplía la diversidad cubierta, pero no produce una muestra representativa de la población ni permite estimar prevalencias. La tarea principal entrega perfil, historia y contexto actual para predecir acciones binarias, duración de vídeo o la siguiente respuesta de atención al cliente. El historial principal cubre septiembre; la prueba, octubre y noviembre; se anuncian 30 acciones de prueba por usuario y 6.000 casos. Se evalúan once modelos. Claude Opus 4.5 obtiene 44,55 en el ranking publicado, seguido por GLM-4.7 con 41,46 y Claude Sonnet 4.5 con 40,49; Qwen3-235B queda en 32,11. Sin embargo, la puntuación global es una media no ponderada de seis componentes heterogéneos: F1 de escenas binarias, 100 menos NMAE para duración y una puntuación textual de Claude-Sonnet-4.5 multiplicada por diez. No se aportan intervalos ni pruebas para las diferencias de ranking, y Claude-Sonnet-4.5 actúa a la vez como competidor, juez textual y anotador de sentimiento/estilo sin validación humana. En memoria, Qwen3-235B obtiene 24,27 con resumen, 21,13 con truncado y 20,38 con RAG; son promedios de escenas distintas y no bastan para generalizar sobre todas las memorias. En 66 historiales de más de 128k tokens, ampliar el contexto no mejora de forma consistente. El artículo interpreta como cadenas causales 180 trayectorias previas a conversiones inferidas por Claude y revisadas manualmente; al ser logs observacionales sin contrafactual ni intervención, son atribuciones retrospectivas, no efectos causales. También compara el drift Jaccard diario de OmniBehavior, 0,6311, con LoCoMo, 0,1698, pero plataforma, tareas y generación difieren y no aíslan una propiedad general de datos reales frente a sintéticos. Sus etiquetas de hyperactivity, utopian bias y personality homogenization dependen de un muestreo equilibrado, anotación de Claude y vectores de tasas de acción; la última mide similitud conductual, no personalidad validada. Hay además inconsistencias de artefacto. Los 2,12 millones sobre 200 usuarios implican 10.600 acciones por usuario, mientras que la media declarada de 8.143 implica 1.628.600, una diferencia no explicada de 491.400. En una auditoría determinista de 20 pares del snapshot público, los JSON ingleses suman 1.324 acciones y los chinos 2.688; solo cinco pares coinciden y la diferencia se concentra en vídeo, 355 frente a 1.719. Esto demuestra deriva en la muestra revisada, no una tasa global. El código exige al menos 60 acciones para construir 30 casos y la versión publicada contiene usuarios auditados por debajo de ese umbral, por lo que no garantiza 200 usuarios ni 6.000 tareas; tampoco incluye configuración reproducible para los once modelos, salidas, juicios o resultados. El dataset es público y no restringido por acceso. Aunque se reemplazan identificadores directos, las marcas temporales, demografía y contenidos raros siguen siendo cuasi-identificadores longitudinales; no se documentan consentimiento, evaluación de privacidad, retirada, contacto de takedown ni rendimiento del proceso de anonimización. El trabajo aporta un recurso industrial rico y código inspeccionable, pero sus resultados deben leerse como un benchmark curado y una prepublicación arXiv v2, no como una simulación representativa, una prueba causal o una validación de personalidad humana.

Research question

To what extent can current LLMs simulate heterogeneous human actions from long, multi-scenario histories, and what structural differences appear between their predictions and real traces?

Method

A benchmark is built with longitudinal logs from Kuaishou. Two hundred users are selected by proximity to K-means centroids defined on demographic traits, activity, interests, and scenario preferences. After cleaning and pseudonymization, binary prediction, video duration, and text generation tasks are prepared. Eleven LLMs receive profile, history, and context; their results are aggregated into a mean of six components. Auxiliary analyses compare context length, truncation/RAG/summarization, interest coverage, trajectories prior to conversions, and similarity of action patterns.

Sample: The article declares 200 users from Kuaishou and 2.12 million interactions between September and November 2025. The selection takes one user close to each of 200 K-means centroids to maximize diversity. The main benchmark announces 30 test actions per user, 6,000 tasks. The auxiliary analyses use 180 conversions and 66 histories longer than 128k tokens. The version audit reviewed 20 English/Chinese pairs from the fixed public snapshot.

Findings

  • Claude Opus 4.5 leads the composite score with 44.55; GLM-4.7 obtains 41.46, Claude Sonnet 4.5 40.49, and GPT-5.2 39.07, but there is no uncertainty to order small differences.
  • The global score is the unweighted mean of six components with distinct units and reliability; the code confirms 100-NMAE for duration and Claude textual by ten.
  • In Qwen3-235B, summarization obtains 24.27, truncation 21.13, and RAG 20.38 in the published heterogeneous average; extending context also does not consistently improve across 66 long histories.
  • Of 180 trajectories attributed to conversions, 81.8% crosses scenarios and more than 60% exceeds three days; the observational design does not identify them as causes.
  • The daily Jaccard drift is 0.6311 in OmniBehavior and 0.1698 in LoCoMo, a descriptive comparison between non-equivalent datasets.
  • The intra/inter-user distance ratio is 0.29 for human traces and 0.66-0.87 for models; it measures similarity of action rates, not psychometric personality.
  • The declared totals are incompatible: 2.12 million/200 equals 10,600 interactions per user, not the reported mean of 8,143.
  • In 20 audited public pairs, English contains 1,324 actions and Chinese 2,688; the gap of 1,364 events corresponds to video and contradicts complete synchronization in that sample.
  • The artifact offers real code and data, but the public version does not guarantee the 200 users/6,000 cases nor reproduce the table of eleven models.

Limitations

  • The sample covers diverse centroids, but does not report base population size, inclusion probabilities, or cluster weights; it is not representative nor suitable for estimating prevalences.
  • The global mean mixes six heterogeneous metrics with unvalidated equal weights; no intervals, bootstrap, tests, or sensitivity to weights or judges are reported.
  • Claude-Sonnet-4.5 competes, judges text, and labels sentiment/style without human calibration, agreement analysis, or alternative judge.
  • Conversion chains are inferred retrospectively over observations; there is no intervention, counterfactual, negative control, or reviewer agreement protocol.
  • The test sampling balances time, domains, and high/low-value actions, so hyperactivity does not directly describe the natural prevalence of the platform.
  • Personality homogenization is operationalized with action rate vectors and not through a validated psychological construct.
  • The English/Chinese snapshot presents action drift and the code excludes users with fewer than 60 records without requiring that 200 remain; splits, outputs, judgments, and configurations for all models are missing.
  • Anonymization does not publish precision/recall, counts, or protocol; temporal traces and rare content retain re-identification risk.
  • Consent, ethical basis, withdrawal, retention, takedown contact, and how the request not to re-identify is reconciled with the CC BY-NC-SA license are not documented.
  • It is an arXiv v2 preprint; the limitations section of the paper focuses on model and memory evolution and omits several of these limits.

What the study does not establish

  • It does not establish that the 200 users represent Kuaishou or that the observed patterns measure the natural frequency of behaviors in the population.
  • It does not prove causal relationships between scenarios and conversions; the chains are plausible explanations of observed sequences.
  • It does not demonstrate that synthetic data in general has less drift than real data, because OmniBehavior and LoCoMo are not equivalent controls.
  • It does not validate personality homogenization as a loss of human personality or utopian bias with independent human annotators.
  • It does not support a stable ranking of eleven models under other weightings, judges, or samples, nor does it quantify the uncertainty of the differences.
  • It does not guarantee that the English release is a complete and synchronized translation of the Chinese or that the published code generates exactly 6,000 tasks.
  • It does not allow reproducing the full table without absent APIs, versions, configurations, outputs, and judge logs.
  • It does not eliminate the risk of re-identification or document consent or effective mechanisms for individuals to withdraw their traces.

Traceability

Scope: Full text

Version: arXiv:2604.08362v2 preprint

Consulted source: https://arxiv.org/pdf/2604.08362v2

Review: Codex 27-page visual full-text, TeX/source, sampling, metric arithmetic, 20-pair bilingual release, privacy, code, artifact and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • Claude Opus 4.5
  • Claude Sonnet 4.5
  • Claude Haiku 4.5
  • Claude Sonnet 4
  • Gemini 3 Flash
  • GPT-5.2
  • GPT-4o
  • GLM-4.7
  • DeepSeek-V3
  • Kimi-K2-Instruct-0905
  • Qwen3-235B-A22B-Instruct-2507
  • Qwen2.5-72B and Qwen3-235B for data processing

Instruments and metrics

  • Per-scene F1 for binary actions
  • Normalized mean absolute error and 100-NMAE for video duration
  • Claude-Sonnet-4.5 textual judge on a 1-10 scale
  • Unweighted six-component overall score
  • Daily Jaccard interest drift
  • 19-dimensional action-rate distance ratio
  • Claude-Sonnet-4.5 sentiment and language-style labels

Data used

  • OmniBehavior Kuaishou longitudinal traces
  • LoCoMo comparison corpus
  • Public Hugging Face OmniBehavior Chinese and English release

Evidence and location

  • Design, sampling, scenarios, cleaning, anonymization, and declared size: Paper, pp. 1-6, sections 1-2 and Figures 1-2
  • Multi-scenario coverage, 180 conversions, and comparison with LoCoMo: Paper, pp. 6-8, section 3 and Figures 3-5
  • Task, eleven models, global score, and main results: Paper, pp. 8-11, section 4 and Table 1
  • Long context, memory, and behavioral bias analysis: Paper, pp. 11-15 and 23-27; Figures 7-10 and Tables 2-6
  • Sampling implementation, 60-action threshold, and metric composition: GitHub commit c00c56dd5cb46e85714ba0ea93ab9241a6fce272; src/data/prepare_experiment_data.py and src/evaluation/metrics.py
  • English/Chinese drift, dates, lengths, and residual identifier risk: Deterministic audit of 20 paired files from Hugging Face dataset commit 28479ffdca40b8657018719b0763e56f2981d254
  • License, public access, and privacy warnings: Hugging Face README and LICENSE, SHA-256 3b8b220e3d2ca89c12dd4a6e5ad66464598969c3a9e442e3beec8404e7e356f0 and 9e502ba75fd0fd30cea5497b438a4bc12f07f02bc932a7a6041fd6a18b8f46a0
  • Comprehensive audit of validity, score, privacy, code, and reproducibility: reports/verification/article-373-omnibehavior-sampling-causal-attribution-composite-score-llm-judge-bilingual-data-drift-privacy-code-and-reproducibility-audit.json