The authors trained a single tool-using LLM policy to search for evidence, reason over it, and assign probabilities to possible outcomes.
Training LLM forecasters across dates improves search and probability estimates
AlphaBet trains one agent to gather time-appropriate evidence and update probabilistic forecasts before events resolve.
Big Tech
Liqin Ye · Rongzhi Zhang · Xin Liu · Zheng Li · Yuan He · Qingyu Yin · +2 more
Amazon Data & Models · Georgia Institute of Technology
Research Digest··2 min read
Ye and colleagues introduce AlphaBet, a reinforcement-learning framework that repeatedly forecasts the same unresolved event at different dates, using only evidence available at each point.
§