Preprint reports autonomous agent gains on production recommender model

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.