Salesforce Releases R&D Agents That Improve Its AI Models
Salesforce, Wednesday, November 11th, 2026
Salesforce AI Research's SFR-AutoR&D agents autonomously run experiments to improve codebases and train production embedding models.
SFR-AutoR&D is a set of agents that run the research loop end to end: given a goal, a system and a compute budget, they form hypotheses, implement methods, run controlled experiments against baselines and keep only changes supported by evidence.
Salesforce says the system autonomously improved SFR-Embedding V3, the retrieval model used in Data Cloud, completing the cycle in about 80 hours, and produced SFR-Embedding V4 from the open source EmbeddingGemma 2 with better quality and latency.
It earlier found gains in pandas, hnswlib and USearch.
The system comprises AutoR&D-Engineer for algorithmic and systems improvements and TrainForge for new training methods, both running on the AutoInfra training backbone.