7B model that keeps updating weights post-deployment beats GPT-5.6 on prediction markets
jiqizhixin · x · 2026-09-30
UIUC, with Tsinghua-founded startup Astraculum, extended the Social World Model framework (ICML 2026) to test whether an LLM's cognitive update can extend beyond training into deployment. Key points:
- The 7B model continuously updates its own weights from live market data and real-time news, letting today's market changes shape tomorrow's judgment
- Polymarket and Kalshi prediction markets serve as the natural testbed, since real-money bets on unsettled questions sharply reflect shifting consensus
- 390 days of real prediction-market data with rolling training and continuous backtesting
- In the same test environment it outperformed GPT-5.6, DeepSeek V4 Pro, and Claude Opus 5 — not because it's bigger, but because it keeps learning from real feedback while the others stay frozen
(Reported by jiqizhixin; details pending verification against the original work.)
More from Research
- Anima Anandkumar to Keynote SC26, Teasing a New Scaling Law Beyond LLMs — AnimaAnandkumar · 2026-09-30
- TypeSafe's Workflow Evals: decomposing policies into workflows beats prompts on accuracy, cost and speed — multiply_matrix · 2026-09-30
- SchmidhuberAI: AI proofs will be like compiler output — individual researchers' value gone in a year — ethanCaballero · 2026-09-30
- TCRvdb: a functionally validated TCR-pMHC database for specificity models — iskander · 2026-09-30
- Task Scheduling as a Bandit Problem: FLD Paper Details Continuous-Time Bandit in Human Motion Learning — breadli428 · 2026-09-30
- RoboPapers teases episode on steering pretrained robot policies with VLMs (VLS) — micoolcho · 2026-09-30