Step-Level Preference Learning for Generative Agents in Social Simulations

Trait induction and control2026arXivApproved editorial review

Authors: Wenchang Gao, Pingyue Sheng, Lanlan Qiu, Yunfei Ma, Jian Zhao, Baicheng Chen, Kangda Wang, Yuyang Tian, Shunqiang Mao, Tianxing He

Keywords: Persona conditioning, Behavioral control, Longitudinal behavior

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

10
Authors
4
Findings
5
Limitations
3
Evidence

Editorial summary

English

Eight engineers use SimPref for four weeks to choose among three GPT-4o candidates in six modules. The study collects 57,239 pairs, trains Qwen2.5-7B/14B and Llama-3.1-8B with SFT and DPO, and evaluates ten held-out events with three three-day episodes and GPT-5.2 and DeepSeek-3.2 judges.

Eight annotators, 30 training events, and 10 held-out events with four to eight agents. Each step has one human label and there is no overlap for estimating agreement. GPT-4o top-1 agreement with humans ranged from 29.3% to 51.4% by module. SFT produced the largest gains; DPO added smaller and uneven gains. For Qwen-14B, behavioral KL to the human reference fell from .610 to .084. Idleness fell from 31% to 22% and coordination rose from 8% to 36% in one reported contrast.

Each step is labeled by one person. The distribution is dominated by two modules. Evaluation depends on two LLM judges. No statistical inference for differences is reported. General-capability retention and transfer to other environments are not tested. It does not validate that agents simulate real people. It does not demonstrate representativeness of eight engineers preferences. It does not isolate DPO as the primary cause of improvement.

Español

Ocho ingenieros usan SimPref durante cuatro semanas para elegir entre tres candidatos generados por GPT-4o en seis módulos. Se reúnen 57.239 pares, se entrenan Qwen2.5-7B/14B y Llama-3.1-8B mediante SFT y DPO, y se evalúan diez eventos retenidos con tres episodios de tres días y jueces GPT-5.2 y DeepSeek-3.2.

Ocho anotadores, 30 eventos de entrenamiento y 10 eventos retenidos con cuatro a ocho agentes. Cada paso tiene una sola etiqueta humana y no existe solapamiento para estimar acuerdo. El top-1 de GPT-4o coincidió con humanos entre 29,3% y 51,4% según módulo. SFT produjo las mayores mejoras; DPO añadió ganancias menores e irregulares. En Qwen-14B, el KL de conducta respecto a referencia humana bajó de .610 a .084. La inactividad descendió de 31% a 22% y la coordinación de 8% a 36% en un contraste reportado.

Cada paso lo etiqueta una sola persona. La distribución está dominada por dos módulos. La evaluación depende de dos jueces LLM. No se reporta inferencia estadística de las diferencias. No se comprueba retención de capacidades generales ni transferencia a otros entornos. No valida que los agentes simulen personas reales. No demuestra representatividad de las preferencias de ocho ingenieros. No aísla DPO como causa principal de la mejora.

Research question

Does human supervision of intermediate decisions improve coordination and fidelity of generative agents across social simulations?

Method

Eight engineers use SimPref for four weeks to choose among three GPT-4o candidates in six modules. The study collects 57,239 pairs, trains Qwen2.5-7B/14B and Llama-3.1-8B with SFT and DPO, and evaluates ten held-out events with three three-day episodes and GPT-5.2 and DeepSeek-3.2 judges.

Sample: Eight annotators, 30 training events, and 10 held-out events with four to eight agents. Each step has one human label and there is no overlap for estimating agreement.

Findings

  • GPT-4o top-1 agreement with humans ranged from 29.3% to 51.4% by module.
  • SFT produced the largest gains; DPO added smaller and uneven gains.
  • For Qwen-14B, behavioral KL to the human reference fell from .610 to .084.
  • Idleness fell from 31% to 22% and coordination rose from 8% to 36% in one reported contrast.

Limitations

  • Each step is labeled by one person.
  • The distribution is dominated by two modules.
  • Evaluation depends on two LLM judges.
  • No statistical inference for differences is reported.
  • General-capability retention and transfer to other environments are not tested.

What the study does not establish

  • It does not validate that agents simulate real people.
  • It does not demonstrate representativeness of eight engineers preferences.
  • It does not isolate DPO as the primary cause of improvement.

Traceability

Scope: Full text

Version: arxiv; 16-page full text reviewed 2026-07-18

Consulted source: https://arxiv.org/abs/2607.14485

Review: Codex full-text and visual 16-page methodological, statistical and claim-boundary review, 2026-07-18

Approval: Codex fidelity pass, 2026-07-18

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4o
  • Qwen2.5-7B
  • Qwen2.5-14B
  • Llama-3.1-8B
  • GPT-5.2
  • DeepSeek-3.2

Instruments and metrics

  • SimPref interface
  • SFT
  • DPO
  • Five trajectory ratings

Data used

  • 57,239 preference pairs
  • 40 social events

Evidence and location

  • Research question, method, results, and discussion: Full text, pp. 1-16, visually reviewed on 18/07/2026
  • Figures, tables, results, and limitations: Primary PDF sha256 a3b980d4c26d2b76bd8f6cf5aac81056e50f6e2cebb2732da58cd591ad2a3cf4; methods, results, limitations, and appendices
  • Editorial decision and claim boundary: Critical record article-408, complete cross-check of 16 pages