# HExA preprint reports agent gains through active experimentation

_Published Monday, October 12, 2026 at 11:09 AM EDT · AI, Science · Latest · Tier 2 — Notable_

Researchers report that their Hierarchical Experimentalist Agents framework enables frontier language models to reach near 100% success on the most challenging tasks in Interphyre, a procedural physics test where those models otherwise fail. The preprint introduces a method that lets agents gather evidence by designing experiments rather than relying only on existing knowledge.

Called HExA, the framework refines experiments, converts experience into reusable skills and uses the resulting evidence to answer questions or act. It requires no external supervision, oracle solutions or offline data. The researchers also report better token efficiency than naive agent and memory baselines, with evaluations spanning reasoning, partially observable games and physics-law discovery.

## Sources

- [cs.LG updates on arXiv.org](https://arxiv.org/abs/2606.29315)

---
Canonical: https://techandbusiness.org/newswire/xJhb3WJY1IJMPVR8ilWzQz
Published: 2026-10-12T15:09:26.411Z
Story chronology: 2026-10-12T04:00:00.000Z
Retrieved: 2026-10-12T17:47:46.864Z
Publisher: Tech & Business (techandbusiness.org)