← All news
Global · AI & Technology
Global·AI & Technology

Salesforce publishes AI research agents that improve code and model training

Salesforce AI Research has published SFR-AutoR&D, an agent system that develops and tests improvements to software and AI models. The company says it is already using the system to improve production models.

AI Hubs United Editorial ·

Salesforce AI Research has published SFR-AutoR&D, a system of agents that proposes hypotheses, implements changes and tests them against baselines within a researcher-defined goal and compute budget. According to Salesforce News, the system is already being used to train and improve production AI models, including SFR-Embedding V3, which supports retrieval in Data Cloud. Work on the open-source EmbeddingGemma 2 model also produced SFR-Embedding v4, which Salesforce reports improved both quality and latency.

The system combines AutoR&D-Engineer, which looks for improvements in existing code, with TrainForge, which develops and evaluates training methods. Both use AutoInfra to run experiments on shared GPU capacity. Salesforce reports that an autonomous research cycle took about 80 hours and improved nDCG@10, a measure of top-ranked retrieval relevance, by 6.13 points across 10 public MTEB tasks. A subsequent session lasting roughly nine hours delivered 36% higher throughput and 17% lower latency, with little quality loss.

Those gains could help enterprise retrieval systems find relevant records more efficiently. Salesforce says researchers retain control over goals, budgets, evaluations and data constraints, and changes are kept only when they outperform the baseline. Models intended for customers must also satisfy production requirements before deployment.

AI Hubs United original short summary of the source linked above. It is not the original article; read the full story at the source.