Salesforce's SFR-AutoR&D lifts embedding model 6.13 points

Salesforce AI Research is pitching SFR-AutoR&D as a research system rather than a single model release. The setup is deliberately constrained: a researcher defines the goal, the compute budget, the evaluation and the data rules, and the agents explore only inside those boundaries. That framing matters because the output is not just a better score but a codebase change backed by benchmark evidence, which is what makes it usable in production rather than only in a paper.

The early results span two kinds of targets. On public open source code, the agents found further gains in heavily optimized libraries such as pandas, hnswlib and USearch, each packaged with benchmark evidence. On Salesforce's own embedding models, the improvement translated into retrieval quality and efficiency: alongside the 6.13-point nDCG@10 gain, a second session focused on runtime came back with 36% more throughput and 17% lower latency with almost no loss in quality. Efficiency gains like these matter because embedding models run across millions of customer records.

The system is presented as complementary to human judgment, not a replacement. Salesforce says the models trained for product use still need to clear production requirements before customers see them, and that work is underway now. Its broader point is that the research loop itself is being automated, with researchers moving toward setting constraints for autonomous systems; a related effort, CodeEvolve, applies evolutionary search to Java and Salesforce Apex code.