AlphaBet · Agentic RL for event forecasting
AlphaBet trains one LLM to acquire evidence, reason, and report probabilities as information evolves.
Reinforcement Learning for LLM Event Forecasting Agents
Georgia Tech Amazon01 The result
02 The idea · longitudinal forecasting training
AlphaBet revisits one question on several dates, each time searching only the news published by then. Each dated score is a training reward for the shared policy.
Every question-date is its own training sample: one resolved event, 3.9 samples on average, no extra labels.
Every scheduled date starts fresh: the agent searches the dated corpus again and reports again. Nothing is carried from one date to the next.
With each forecast the agent also writes a belief notebook: its probabilities, supporting evidence, open questions. Only the notebook is carried to the next date; everything else starts fresh.
03 Two propositions
Proposition 1 · agentic forecast properness
Zero regret exactly when p = qH = q at every date. One outcome scores the sum of both gaps.
Proposition 2 · GRPO without std
mean-centered · AlphaBet
A = R − mean(R)
Same direction, scaled by 1 − 1/G.
standardized · vanilla GRPO
A = (R − mean(R)) / (σ + ε)
As G grows, the target tilts to q̃H unless every σk is equal.
04 Analysis
Same 1,028 events; one forecast date per event, or all of them. The gap widens farther from settlement.
Freeze the articles each agent retrieved and hand them to tool-free readers. Swapping the reader moves Brier about four times as much as swapping the evidence, and AlphaBet’s evidence helps every reader.
The same training with and without GRPO’s standard-deviation step. With it, both proper scores come out worse and the policy grows over-confident; accuracy remains comparable.
05 In short
AlphaBet is an outcome-based RL recipe for agentic forecasting: one policy, many dated information states, one proper-score reward per date, and GRPO without standard-deviation normalization, so the reward keeps its scale.
A resolved event properly scores each pre-resolution forecast, turning 1,028 events into 3,965 training states.
The gains come from better evidence and better probabilities: AlphaBet’s retrieved articles help every reader, and AlphaBet reads the same articles better.
Standard-deviation normalization tilts the proper-score target and over-sharpens forecasts; mean-centering alone keeps the incentive honest.
The paper
Reinforcement Learning for LLM Event Forecasting Agents
1Georgia Institute of Technology · 2Amazon · †Work done during an internship at Amazon · Corresponding author(s): {liqiny, chaozhang}@gatech.edu
Cite this work
@misc{ye2026alphabetreinforcementlearningllm,
title={AlphaBet: Reinforcement Learning
for LLM Event Forecasting Agents},
author={Liqin Ye and Rongzhi Zhang and
Xin Liu and Zheng Li and Yuan He
and Qingyu Yin and Priyanka Nigam
and Chao Zhang},
year={2026},
eprint={2610.22684},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2610.22684},
}