# SynAct preprint reports agent accuracy gains from targeted expert feedback

_Published Monday, October 12, 2026 at 10:10 AM EDT · AI, Science · Latest · Tier 2 — Notable_

Researchers report that SynAct, a framework for directing expert feedback to consequential agent decisions, improved held-out accuracy by +11.7pp on average across four benchmarks and four models. The preprint reports improvements in all 16 evaluated settings.

SynAct compares repeated workflow attempts to identify decisions associated with different final outcomes, requests expert corrections at those points, and converts the corrections into reusable rules. Agents can then apply those rules without expert access during deployment.

The same signal also guided routing during inference, outperforming budget-matched alternatives in 15 of 16 settings. The findings concern the evaluated benchmarks and models.

## Sources

- [cs.LG updates on arXiv.org](https://arxiv.org/abs/2610.12651)

---
Canonical: https://techandbusiness.org/newswire/tPNndnJiz8XJZxkBGRFT4d
Published: 2026-10-12T14:10:10.221Z
Story chronology: 2026-10-12T04:00:00.000Z
Retrieved: 2026-10-12T16:38:48.804Z
Publisher: Tech & Business (techandbusiness.org)