Salesforce releases SFR-AutoR&D agents that train and improve AI models
Salesforce AI Research published SFR-AutoR&D, a set of agents that runs the research loop itself: forming hypotheses, implementing methods, running controlled experiments and keeping only changes that beat baselines. Salesforce says the system is now used to train production models. It improved SFR-Embedding V3, used for retrieval in Data Cloud, and produced SFR-Embedding v4 from the open-source EmbeddingGemma 2 with better quality and latency. AutoR&D-Engineer and TrainForge run on AutoInfra, shared GPU infrastructure. On 10 public MTEB retrieval tasks Salesforce reports a 6.13-point nDCG@10 gain; a second run raised throughput 36% and cut latency 17%. A cycle that historically took weeks finished in about 80 hours.
Why it matters
Embedding models determine what an AI agent retrieves from enterprise data, so a faster research loop can improve the grounding behind answers and reduce inference costs. Salesforce says the approach is already applied to production models, though customer-facing deployment still requires trust and safety checks.
Sources
- PrimarySalesforce
Written by AIToolMaps from the source above and checked against it sentence by sentence.