Socio-Foundation: A Model for Generalizable Individual Behavior Simulation via Hierarchical Capability Distillation

2026-10-06

AI Summary

This study addresses the limitations of general-purpose large language models in individual behavior simulation, specifically role flattening and insufficient fine-tuning generalizability. To this end, it proposes FONTS, a five-dimensional capability framework, and constructs a standardized corpus comprising tens of millions of instances. Methodologically, an expert-decoupled multi-teacher online policy distillation paradigm is designed, integrated with the DAPO algorithm to implement three-stage hierarchical capability training. Experimental results demonstrate that the proposed approach outperforms baselines by 11.0 points, achieving performance comparable to state-of-the-art frontier models while significantly enhancing both the generalizability and naturalness of behavior simulation.

Abstract

Simulating individual behavior requires large language models (LLMs) to preserve persona traits while adapting to dynamic social contexts. However, general-purpose LLMs often flatten distinct personas, while task-specific tuning suffers from fragmentation and generalization. To overcome these challenges, we organize individual simulation into the \textbf{FONTS Taxonomy}, comprising five complementary capability dimensions: \emph{persona fidelity} (\textbf{F}), \emph{outcome realization} (\textbf{O}), \emph{behavioral naturalness} (\textbf{N}), \emph{trajectory coherence} (\textbf{T}), and \emph{social grounding} (\textbf{S}). Grounded in this taxonomy, we curate a standardized training corpus library of approximately 10 million instances across 14 representative datasets and present \textbf{Socio-Foundation}. Socio-Foundation decouples specialization from integration via a three-stage pipeline: learning task experts via DAPO, consolidating them into capability experts via off-policy distillation, and unifying them via multi-teacher on-policy distillation (MOPD). We also establish \textbf{IndiEval}, consolidating 29 metrics across the FONTS dimensions. Experiments show that Socio-Foundation outperforms its \textit{Qwen3-8B} base by 11.0 points and approaches frontier models such as \textit{GLM-5.2}, with ablations and out-of-distribution evaluations further demonstrating the effectiveness and generalization of our model.

Problem

• Individual Behavior Simulation
• Large Language Models
• Persona Fidelity
• Generalization
• Task-specific Tuning

Innovation

• Hierarchical Capability Distillation
• Individual Behavior Simulation
• FONTS Taxonomy
• Multi-teacher On-policy Distillation
• IndiEval