macro2mind Trains LLMs on Prediction Markets to Simulate Individuals, +15.5 Points Zero-Shot
youjiaxuan · x · 2026-10-08
Researchers released macro2mind, a method that trains LLMs to reason about individual beliefs and behaviors using prediction market dynamics, without individual-level supervision.
- Method: Uses GRPO to train a language model on market signals. A social behavioral decomposition makes behavioral reasoning an explicit forecasting step: the model infers representative groups of market participants, predicts how each interprets news and updates beliefs, reasons about their interactions, and aggregates responses into a price.
- Training: A hindsight-regret curriculum with difficulty-aware sampling focuses on transitions where hindsight-identified groups substantially improve the forecast while remaining learnable for the current policy.
- Results: Achieves state-of-the-art directional accuracy and correlation on Polymarket in SWM-Bench. Trained only on market data, it transfers zero-shot to four user-simulation benchmarks (Humanual, OvertonBench, PRISM, CAD), with +15.5 points in unseen-user simulation accuracy.
The paper is by Yining Zhao and 7 co-authors, with Jiaxuan You as corresponding author.
More from Research
- OpenAI math repo formalizes ~42% of top-line results, withdraws 3 papers over sign error — danintheory · 2026-10-08
- OpenAI cracks k-median open problem that a research group worked on for over a year — CSProfKGD · 2026-10-08
- Building a computer out of fluids: simulating OpenAI's Navier–Stokes Turing machine — kfountou · 2026-10-08
- D-OPCD distills agent harness gains into diffusion weights, lifting T2I scores 60.52 to 65.09 — Wenxuan Wang · 2026-10-08
- QuadTok: quadtree visual tokenizer saves 10% tokens, hits 2.08 gFID on ImageNet 256 — Yucheng Mao · 2026-10-08
- PhysEvo: recursive self-improvement around a frozen model hits 62% on RoboDojo, 84% on real robot — Wenqing Tian · 2026-10-08