Salesforce says research agents improved retrieval quality and model speed

Salesforce says its SFR-AutoR&D agents are being used to improve production AI models, including the embedding model used for retrieval in Data Cloud. The agents formulate hypotheses, modify code and run controlled experiments within goals and compute budgets set by researchers. In evaluations across 10 public retrieval tasks, Salesforce reported a 6.13-point improvement in nDCG@10, which measures the relevance and ordering of the top 10 results. A second session delivered 36% more throughput and 17% lower latency with almost no quality loss. Models intended for customers must still clear production requirements; researchers retain control over evaluation and data constraints.