Salesforce AutoR&D Agents Improved Its Embeddings

Salesforce on 11 November 2026 said it has published SFR-AutoR&D, a set of research agents built to run a model-improvement loop with less day-to-day human steering. In a company newsroom post, Salesforce AI Research wrote that, given a goal, a system to work on and a compute budget, the agents “formulate hypotheses, implement methods, and run controlled experiments against baselines,” and “retain improvements to the codebase only when supported by evidence.”

The worked example is an embedding model. The system, Salesforce says, “autonomously improved SFR-Embedding V3, the model Salesforce uses for retrieval in Data Cloud.” The team then “pointed SFR-AutoR&D at the open source model EmbeddingGemma 2, and it produced a measurably better model in both quality and latency, SFR-Embedding v4.” Those are Salesforce’s results, from its own evaluation, not an independent leaderboard ranking.

Two agents do the work. AutoR&D-Engineer “searches for algorithmic and systems improvements inside existing code.” TrainForge “explores new training methods, runs them, and evaluates candidates against held-out benchmarks that the system never trains on.” Both run on AutoInfra, “the training backbone that carries their experiments across shared GPU capacity.” Salesforce says the system had already posted earlier gains on three heavily tuned open-source libraries — pandas, hnswlib and USearch — “each packaged with the benchmark evidence.”

Salesforce’s own scores, in two checkpoints

The newsroom post reports one headline quality figure. “Evaluated against 10 public retrieval tasks in MTEB,” it says, SFR-AutoR&D raised the model’s nDCG@10 score by 6.13 points. The quality cycle, it adds, completed “in about 80 hours.” A second session, “roughly nine hours,” aimed at how fast and cheaply the model runs, “came back with 36% more throughput and 17% lower latency, with almost no loss in quality.”

The SFR-AutoR&D embeddings page splits that quality claim into two checkpoints, both of 272.6 million parameters, and it is careful about what the numbers are. A research version, sfr-embedding-rnd_R, with 2,048-dimensional embeddings, is shown moving from a 51.07 baseline to 57.21, a gain of 6.13 points in mean retrieval nDCG@10 across those 10 tasks. A “production supported version,” sfr-embedding-rnd_P, with 768-dimensional output, is shown at 55.76, a gain of 4.69 points over the same style of baseline. The page says: “These local measurements compare evaluated checkpoints; they are not official leaderboard ranks.” The 6.13-point figure in the newsroom post is the research checkpoint; the production-supported run is the smaller 4.69-point lift.

None of this is a customer-facing model release. Salesforce writes that “a model that performs well in evaluation still has to clear production requirements before any customer sees it, and that work is underway now.” SFR-Embedding v4 is not product general availability. The “production supported” label on the research site describes a checkpoint and a data policy, not a model already serving customers.

Silvio Savarese, executive vice president and chief scientist, said: “The scientific method still governs, and the loop runs faster.” Shafiq Joty, senior research director at Salesforce AI Research, framed the agents as systems that “independently formulate hypotheses, develop novel methods, and validate them through controlled experiments on held-out evaluations.” A related Salesforce research effort, CodeEvolve, which the newsroom says uses evolutionary search to improve Java and Salesforce Apex, is promised in a later blog and is not part of these embedding scores.

A research loop is not a shipped model

The result to take seriously is also the one Salesforce qualifies. A pair of agents, on the company’s own account, can run a retrieval-training loop in a few days and show a higher nDCG@10, then spend about nine hours chasing throughput and latency. The scores are local measurements on 10 MTEB retrieval tasks, and the better of the two gains belongs to a wider, research checkpoint rather than the production-supported one. The model those agents improved still has to pass Salesforce’s production requirements before any customer retrieval stack uses it.

That is a tighter claim than “AI is now improving Salesforce’s production models on its own.” It is still a notable research system, especially because it was pointed at EmbeddingGemma 2 and at a model used inside Data Cloud, and because Salesforce is already buying more of the agent stack around customer conversations. The agreed acquisition of Cresta was about voice agents in the call centre, not about embeddings, but it is the same company pushing agents further into products customers already run. Spending caps and evaluation checks announced for Agent Fabric are the operational counterpart to the line that a stronger checkpoint still has to clear production requirements. The $1 billion Spain investment plan, with Madrid as an EMEA South hub, is the geographic expansion sitting behind that research organisation; it does not change the status of SFR-Embedding v4.

The facts to watch are a customer-facing Data Cloud announcement that names v4, and any official MTEB leaderboard entry that matches the local 57.21 or 55.76 figures. Until then, SFR-AutoR&D is a research system Salesforce says improved its own embeddings, and those embeddings are not yet a generally available product.