PsyAgent: Constructing Human-like Agents Based on Psychological Modeling and Contextual Interaction

TL;DR

This work presents PsyAgent, which couples a Big Five trait prior with Bourdieu's cognitive-social co-structure, and offers a precise, data-efficient architecture for personality-grounded agents.

Abstract

Human-like agents require modeling how dispositions interact with social structure. We present PsyAgent, which couples a Big Five trait prior with Bourdieu's cognitive-social co-structure. PsyAgent comprises: (i) Individual Structure (IS), a machine-usable profile encoding traits and facets, cognitive style, values, cultural and educational capital, and salient life episodes; and (ii) Multi-Scenario Contexting (MSC), role-relationship-norm frames spanning eight arenas (work, family, friendship, strangers and civic life, solitude and self-regulation, romance, learning, and public expression). At inference, fixed structured prompts bind the active scenario to the agent profile, yielding behavior that is stable yet context-sensitive. We instantiate IS and MSC to synthesize supervision (role-play dialogues, decision probes, feedback trajectories) and then fine-tune a small LLM. The resulting model produces consistent, identifiable persona-aligned behaviors for specified Big Five configurations and matches or exceeds several larger untuned LLMs and other untuned baselines on our metrics: persona consistency, contextual appropriateness, style matching, trait identifiability, and long-horizon stability. Ablations show IS chiefly improves trait fidelity and stylistic stability, while MSC drives norm awareness and decision fit; both are necessary for cross-scenario performance. PsyAgent offers a precise, data-efficient architecture for personality-grounded agents.

Figures (22)

• Figure 1: Multi-Scenario Contexting (MSC). Eight arenas—Working, Family, Friendship, Strangers, Solitary, Romantic, Learning, Public—each with representative subskills.
• Figure 2: Individual Structure (IS) schema. Four domains—(1) Educational Trajectory: stages, specialization, performance, pedagogy, mentors, transitions; (2) Life Experience: origins/mobility, roles, critical events, travel, social style, interests; (3) Socioeconomic Context: family, home culture, networks, class identity, mobility; (4) Cultural Capital: embodied, objectified, institutional, taste, media habits, cultural consumption.
• Figure 3: Overview of the PsyAgent pipeline. Individual Structure (IS) profiles are crossed with Multi-Scenario Contexting (MSC) frames and a target Big Five vector to form a unified persona prompt with control tags ($<O_{\cdot}>$, $\langle$SCENE=$\cdot\rangle$, $\langle$INSTR$\rangle$, $\langle$RESP$\rangle$). In SFT, a frozen base LLM with LoRA adapters is trained by cross-entropy only on tokens after$\langle$RESP$\rangle$; in DPO, chosen–rejected pairs for the same prompt further optimize the adapters against a frozen reference (optionally with 4-bit QLoRA). The resulting small, adapterized model generates high-quality, scene-conditioned outputs that faithfully express the specified Big Five persona.
• Figure 4: Individual Structure (IS) schema. Full, machine-usable breakdown of Educational Trajectory, Life Experience, Socioeconomic Context, and Cultural Capital. The figure is normative: it defines field granularity and serialization order used throughout data generation and analysis.
• Figure 5: Multi-Scenario Contexting (MSC) schema. Eight arenas with role–relationship–norm scaffolds and representative subskills. Frames are authored once and reused across targets, enabling context-rich conditioning at train and test time.