Salesforce Releases R&D Agents That Improve Its AI Models

Salesforce AI Research unveiled SFR-AutoR&D, a self-improving system that autonomously trains and enhances embedding models used in enterprise AI deployments. The agents formulate hypotheses, implement methods, and conduct controlled experiments against baselines, retaining only improvements supported by evidence and safety checks. The system is already operational, having autonomously improved Salesforce’s production model SFR-Embedding V3 for Data Cloud retrieval tasks.

The research team applied SFR-AutoR&D to the open source model EmbeddingGemma 2, producing a measurably better model, SFR-Embedding v4, with improvements both in quality and latency. The system also enhanced widely used open source libraries like pandas, hnswlib, and USearch, which power data manipulation and similarity search in enterprise AI workflows. Each improvement was packaged with benchmark evidence demonstrating measurable gains over existing baselines.

Embedding models convert text into numerical representations to retrieve relevant records from large datasets, a critical function for enterprise AI agents. Historically, improving these models required weeks of human-led experimentation, but SFR-AutoR&D completed the research loop in about 80 hours by autonomously testing hypotheses and validating results. The system’s agents operate within predefined goals, compute budgets, and ethical constraints to ensure trustworthy outcomes for production use.

Salesforce’s approach emphasizes rigorous experimental validation, with improvements retained only if they demonstrate measurable gains over baselines. The company treats ethical and legal review as part of the problem definition, ensuring models meet production requirements before deployment. Researchers retain control by setting goals and constraints, while the system explores solutions autonomously within those boundaries.