Socio-Foundation Targets Generalizable Individual Behavior Simulation

Summary

The paper presents Socio-Foundation, a language model designed to simulate individual behavior while preserving persona traits and adapting to changing social contexts. It argues that general-purpose LLMs often make different personas too similar, while task-specific tuning can fragment capabilities and generalize poorly. To structure the problem, the authors propose the FONTS taxonomy: persona fidelity, outcome realization, behavioral naturalness, trajectory coherence, and social grounding. They assemble a training corpus of about 10 million instances from 14 datasets. The model is trained through three stages: task experts are learned with DAPO, consolidated into capability experts through off-policy distillation, and unified with multi-teacher on-policy distillation. The paper also introduces IndiEval, which combines 29 metrics across the five FONTS dimensions. In experiments, Socio-Foundation scores 11.0 points higher than its Qwen3-8B base and approaches the performance of GLM-5.2. Ablation and out-of-distribution evaluations are reported as further evidence that the design improves capability integration and generalization.