# AlphaBet preprint reports improved probability forecasts through agent training

_Published Monday, October 19, 2026 at 12:14 AM EDT · AI, Science · Latest · Tier 2 — Notable_

Researchers report that AlphaBet, a reinforcement learning framework, improved GLM-4.5-Air's event probability forecasts on two benchmarks. Brier scores, which measure probability forecast error, fell from 0.731 to 0.626 on Forecast-Dojo and from 0.612 to 0.528 on FutureX.

The framework trains one model to search for evidence, reason and report probabilities at multiple dates before an event resolves. Each forecast is scored against the eventual outcome. Controlled experiments attributed gains to both evidence gathering and probability reporting. The preprint also found that a normalization step in Group Relative Policy Optimization degraded probability scoring while leaving top-outcome accuracy largely unchanged.

## Sources

- [cs.LG updates on arXiv.org](https://arxiv.org/abs/2610.22684)

---
Canonical: https://techandbusiness.org/newswire/wzHpLQRXolKW8lLkdjALZn
Published: 2026-10-19T04:14:45.033Z
Story chronology: 2026-10-19T04:00:00.000Z
Retrieved: 2026-10-19T06:53:07.415Z
Publisher: Tech & Business (techandbusiness.org)