# RSIHub preprint reports agent task gains with uneven transfer

_Published Monday, October 26, 2026 at 6:40 PM EDT · AI, Science · Latest · Tier 2 — Notable_

Researchers report a largest observed gain of 13.5 percentage points in overall first-attempt task success in evaluations using RSIHub, an open-source framework for methods that improve AI agents through execution feedback. The preprint says method rankings varied across settings and optimization gains did not always transfer to held-out tasks.

RSIHub lets researchers combine reusable operations to modify prompts, skills, tools and orchestration while keeping evaluation outside those changes. The researchers implemented AHE, Hyperagents, A-Evolve and GEPA without changing the framework core, then evaluated them with MiniSWE and Codex on Terminal Bench 2 and τ³ Banking under fixed model and evaluator configurations.

## Sources

- [cs.AI updates on arXiv.org](https://arxiv.org/abs/2610.32253)

---
Canonical: https://techandbusiness.org/newswire/tA23Qo2p0IrZjxDv8nXxJk
Published: 2026-10-26T22:40:46.503Z
Story chronology: 2026-10-26T04:00:00.000Z
Retrieved: 2026-10-27T01:15:13.179Z
Publisher: Tech & Business (techandbusiness.org)