Salesforce’s SFR-AutoR&D Automates Model Research and Validation
What happened
On November 11, 2026, Salesforce launched SFR-AutoR&D, an automated system designed to speed up the process of developing and testing AI models. The system produced a new version of Salesforce’s embedding model, SFR-Embedding v4, in about 80 hours, a task that typically takes several weeks for human researchers. SFR-AutoR&D runs experiments by generating and testing new code and training methods without human intervention, though researchers still set the goals and constraints. The new model is used to improve how Salesforce retrieves customer data for its Agentforce platform.
Key details
- SFR-AutoR&D completed SFR-Embedding v4 in around 80 hours.
- The system divides work into four stages: hypothesis discovery, candidate building, evaluation, and learning from results.
- On a public benchmark, the research checkpoint scored 57.21 versus the baseline’s 51.07, a 6.13-point improvement.
- A separate session improved serving efficiency by 36% higher throughput and 17% lower latency.
- Engineering changes included faster pandas reductions and increased vector search speed in hnswlib and USearch libraries.
Why it matters
Salesforce customers and users of its Agentforce platform could see faster and more accurate retrieval of customer data, which affects how well AI-powered responses are grounded in relevant information. The automation of research and validation processes reduces the time and resources needed to develop new AI models, potentially lowering costs and increasing the pace of updates. However, the new model’s full rollout depends on passing additional trust, safety, and production reviews.