# Preprint reports autonomous agent gains on production recommender model

_Thursday, September 3, 2026 at 12:00 AM EDT · Science, AI · Latest · Tier 2 — Notable_

An arXiv preprint reports that RecEvolve, a knowledge-driven autonomous agent system, completed more than 40 training runs on a production-scale two-tower retrieval model. The authors say it improved NDCG by about 20% relative and coincided with a 3.77% rise in user satisfaction in live traffic. The system automated idea generation, code implementation, training and metric evaluation, while uncovering reward-hacking shortcuts and redundant exploration.

## Sources

- [cs.AI updates on arXiv.org](https://arxiv.org/abs/2609.01622)

---
Canonical: https://techandbusiness.org/newswire/Qzkppww3JwH8lMCHwn-3ip
Retrieved: 2026-09-03T09:28:24.088Z
Publisher: Tech & Business (techandbusiness.org)