Building Social World Models with Large Language Models

Summary

Defines a Social World Model as P(s_{t+1}|s_t,e_t) over market-implied beliefs, with a prior news attributor, a frozen hindsight posterior, and an event-conditioned world-model head trained by posterior-guided ELBO-style distillation. Releases SWM-Bench from Kalshi and Polymarket (Dec 2022–Jan 2026): 12,789 volatility-filtered (s_t, E_t, s_{t+1}) triples over 3,248 markets. Qwen3-8B SWM (prior) reports attributed directional accuracy 0.845 on Kalshi and 0.685 on Polymarket, beating time-series and prompted GPT-5.5 on Kalshi direction; Polymarket magnitude still trails frontier LLMs. Code https://github.com/ulab-uiuc/social-world-model; data https://huggingface.co/datasets/ulab-ai/swm-bench.

Keywords

swm · swm-bench · social-world-model · prediction-markets · kalshi · polymarket · attribution · uiuc

Topics

social world models, prediction markets, LLM forecasting

Research notes

• Primary: arxiv abs (cs.LG; ICML 2026 keywords). Correspondence haofeiy2@illinois.edu. Code https://github.com/ulab-uiuc/social-world-model (MIT in paper appendix; 26 stars at check). Data https://huggingface.co/datasets/ulab-ai/swm-bench (223 dl, 1 like, lastModified 2026-06-09; HF card license apache-2.0; paper appendix B.1 says dataset CC BY-NC-ND 4.0 — conflict noted, paper statement recorded on datasets_local). ArXiv license widget not visible in converted abs HTML, so paper license left blank. Discord posted HF dataset + abs. Substantial new corpus; datasets_local row added.