The Tsinghua-based startup team, Astraculum, has teamed up with scholars from the University of Illinois at Urbana-Champaign (UIUC) to introduce the Social World Model (SWM) framework. Their objective is to investigate whether large language models (LLMs) can engage in continual learning and update their cognitive abilities post-deployment. Utilizing prediction markets like Polymarket and Kalshi as their experimental arenas, the framework reframes the process of updating social cognition as a state transition challenge. This means harnessing current market sentiments and breaking news to forecast subsequent shifts in public opinion.

To realize this ambition, the team implemented a self-distillation mechanism akin to SDFT, establishing a teacher-student closed loop through post-hoc reasoning. This approach empowers the model to grasp the logical reasoning chain connecting news events to belief changes, rather than simply conducting numerical fitting. Following 390 days of rolling training and backtesting with real-world market data, this continually updated 7B model outshone state-of-the-art static large models, including GPT-5.6 and Claude Opus 5, in prediction markets. It sustained stable performance throughout the year, with no decline in reasoning prowess.

Furthermore, the team developed the TrajOps engine, specifically tailored for continuous learning. This engine offers collection, inference, and training interfaces to align with the SDFT scheme and facilitate the model's ongoing operation in real markets.