Salesforce said on 11 November that its AI Research team had released SFR-AutoR&D, a system of agents that runs experiments to improve AI models and codebases. The team used it to produce SFR-Embedding v4 from the open-source model EmbeddingGemma 2.
Embedding models turn text into numbers that help a system retrieve relevant records. Salesforce says one agent, AutoR&D-Engineer, searches existing code for improvements, while another, TrainForge, tests training methods against benchmarks kept separate from training. Their experiments run on AutoInfra, which shares GPU capacity between them.
Finding a relevant record among many could become more reliable if an improved model ranks the right result nearer the top. That would matter when a question is answered using stored information, though the retrieval result would still depend on the model used.
Salesforce says the improvement cycle took about 80 hours. Across 10 public retrieval tasks in MTEB, it reports a 6.13-point rise in nDCG@10, a measure that gives greater weight to relevant results near the top. A further session of roughly nine hours produced 36% more throughput and 17% lower latency, with almost no loss in quality, according to the company.
Researchers set the goals, budget and data constraints. Salesforce says models intended for product use must also meet production requirements, and that work is underway.
.jpg){kind=link}