Tsinghua-born startup Astraculum and UIUC use a 7B model with deployment-time self-distillation to beat GPT-5.6 and Opus 5 on prediction markets

Tsinghua-born startup Astraculum and the University of Illinois Urbana-Champaign describe a Social World Model framework that treats updates to crowd beliefs on Polymarket and Kalshi as state transitions of a world model. Using an SDFT-like self-distillation mechanism, the team folds real market feedback back into parameters during deployment. On 390 days of real market data, with monthly deploy–feedback–train–update cycles, a 7B model was the only model to beat a pure price baseline, and the team reports it outperformed GPT-5.6, DeepSeek V4 Pro and Claude Opus 5 in the same test environment, with twelve months free of large drawdowns or reasoning degradation. They also built TrajOps, a continuous-learning MLOps engine to support the pipeline.

Original sources (Chinese)