RESEARCH RELEASE · OCT 2026 ↗

Socio-Foundation

Every person is different.
Simulation should be too.

Preserving individuality in a changing world.
Across diverse social contexts. Powered by an 8B foundation model for individual simulation.

Fudan University · Shanghai Innovation Institute · Tencent

A SMALL WORLD, FULL OF DIFFERENCES.WORKFRIENDSCHOICESA shared worldDistinct individual trajectoriesConceptual animation · not model-generated simulation
8B

Unified simulation model

5

FONTS capability dimensions

+11.0

IndiEval average score gain ①

18.9M

Released SocioBank records ②

① Compared with Qwen3-8B: 50.7 → 61.7. ② Released corpus size; approximately 10M records were selected for training.

APPLICATIONS / INDIVIDUAL SIMULATION

Let your model bring diverse individuals to life.

A learner with doubts. A customer with a deadline. A buyer with firm limits.

Same situation. Different people.Explore all applications ↗

01 / INDIVIDUALS IN CONTEXT

One shared world.
Different choices.

Behavior reflects both a person and their situation.
Switch scenes to explore different responses.

Illustrative interaction
THE SITUATION

02 / THE FONTS TAXONOMY

What does it take to simulate a person?

Five complementary capabilities connect diverse tasks.
A shared structure for training and evaluation.

03 / HIERARCHICAL CAPABILITY DISTILLATION

Learn specialized skills.
Build shared capabilities.

Task experts → capability experts → a unified model.
Decoupling specialization from integration.

One model at inference time.

Deploy Socio-Foundation without loading task or capability experts.

04 / EVIDENCE, NOT JUST PROMISE

A compact model.
A broader behavioral repertoire.

On the IndiEval evaluation reported in the paper,
the five-dimension average rises from 50.7 to 61.7.

0–100 · higher is better

INDIEVAL / LEADERBOARD

Explore the IndiEval leaderboard

Compare all models across five capabilities and 29 metrics.

All compared models and scoring protocol +

Results reported in Table 1. Avg. averages metrics within each dataset, then datasets within each capability, and finally the five capabilities equally. These scores measure individual-simulation performance.

THE PAPER

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

Liang Wang* · Wenxuan Xie* · Xinyi Mou · Yixin Luo · Zhongyu Wei†

Fudan University · Shanghai Innovation Institute · Tencent
* Equal contribution † Corresponding author

View citation

Paper figure

Socio-Foundation hierarchical training frameworkView in the full paper ↗