Salesforce says its SFR-AutoR&D agents can run research experiments and improve AI models, with one model-improvement cycle completed in about 80 hours. The company reports gains in retrieval quality, throughput and latency, offering a concrete example of AI doing part of the work of developing AI.
Salesforce Watch analysis
What happened
Salesforce AI Research describes SFR-AutoR&D as a system that proposes hypotheses, changes code or training methods, runs experiments and keeps improvements only when they beat established baselines. Researchers set the goal, compute budget, evaluation and data constraints.
The company says the system improved SFR-Embedding V3, used for retrieval in Data Cloud, then produced SFR-Embedding v4 from work on the open-source EmbeddingGemma 2 model. Salesforce reports that v4 scored 6.13 points higher on nDCG@10 across ten public retrieval tasks. In a separate optimisation run, it says throughput rose 36% and latency fell 17%, with almost no loss in quality. Salesforce also says the agents have found gains in pandas, hnswlib and USearch. Read Salesforce’s account.
Why it matters
The notable step is not simply getting a model to write code. SFR-AutoR&D is being used to test changes against baselines and held-out evaluations, with the aim of producing measurable improvements. Salesforce says an embedding-model cycle that could take researchers weeks took about 80 hours here.
That could make experimentation faster and help teams improve the retrieval systems behind AI products. But the reported results are Salesforce’s own, and a benchmark gain is not the same as proof of better outcomes for customers. The company says production requirements still need to be met before a model reaches customers.
Our read
This is a substantial example of AI-assisted research with numbers attached, not yet a reason to declare the scientists redundant. The useful idea is a bounded loop: people define the problem and constraints; agents explore possible changes; experiments decide what survives. The test now is whether those gains hold up outside the company’s reported evaluations and translate into dependable products.
What to watch
- Whether Salesforce publishes further evaluation details or independent comparisons.
- Whether the reported quality and efficiency gains persist in production use.
- How researchers set limits and check proposed improvements before deployment.
- Whether similar systems deliver repeatable gains across other models and codebases.
Discussion spark: If an AI research agent can show a measurable gain against a baseline, what further evidence should be required before its change reaches a production model?
Sources and evidence
- Salesforce Releases R&D Agents That Improve Its AI Models (11 November 2026, 16:00 UTC)
Independent WittyWires tracker for public updates about Salesforce. Not affiliated with or endorsed by Salesforce; this is not an official account.