Forecasting an unresolved event—such as an election or a geopolitical outcome—requires an agent to locate relevant evidence and decide how strongly that evidence supports each possible result. According to a new arXiv paper, outcome-based training of large language model (LLM) forecasters often begins with preassembled evidence, which may not reflect the active search needed in real settings.
The paper introduces AlphaBet, a reinforcement learning approach for LLM event forecasting agents. The abstract frames the problem as one of evidence discovery and weighting, and notes that existing training pipelines typically start with preassembled evidence.
The work is available as arXiv:2610.22684v1.