Salesforce says agents built an embedding model in about 80 hours

Salesforce introduced SFR-AutoR&D on November 11, 2026, saying its agents produced SFR-Embedding v4 in about 80 hours, rather than the several weeks human experimentation often requires. The company introduced the system as an autonomous research loop that can change codebases and develop and test new training methods for AI models.

Before researchers pointed the system at the open-source EmbeddingGemma 2 baseline to create v4, SFR-AutoR&D had already improved SFR-Embedding V3. Salesforce uses V3 for retrieval in Data Cloud, making the project more than a standalone research demonstration. Silvio Savarese, Salesforce's executive vice president and chief scientist, said: "The scientific method still governs, and the loop runs faster."

According to Salesforce, embedding models convert text into numerical representations so systems can retrieve records, documents or case histories. In the company's architecture, Data Cloud retrieval supplies customer information that grounds Agentforce answers, so retrieval quality affects the evidence behind each response.

How much autonomy does SFR-AutoR&D actually have?

Salesforce says SFR-AutoR&D divides work into four stages: discover a testable hypothesis; build a candidate and baseline comparison; evaluate correctness or model quality; and learn from successful and failed experiments. A candidate that cannot demonstrate a verified gain leaves the baseline unchanged.

• AutoR&D-Engineer accepts a repository, target and constraints, then produces a candidate patch with benchmark evidence.
• TrainForge starts with a baseline model, task, evaluation and compute budget, then produces the best verified checkpoint and its evaluation record.
• Codex writes implementations from an empty workspace, while checks verify that running code matches the proposed idea.

According to Salesforce, TrainForge can generate new training strategies instead of choosing only from a fixed recipe list. Promising candidates advance from a fast LoRA test, which updates a compact set of trainable parameters while retaining the base model, to a short full-parameter run and then full training. Frozen held-out datasets remain reserved for evaluation.

Both agents run on AutoInfra, Salesforce's shared training and inference infrastructure. It assigns GPU workers to parallel experiments and persists requests, progress and results so work can resume after gateway interruptions. AutoInfra also separates algorithm, training-service and sandbox environments to protect system integrity and guard against reward hacking.

Takeaway: The system automates experiments, while researchers retain control over goals, budgets, evaluations and data constraints.

What do the benchmark and efficiency gains show?

On a public embedding evaluation, Salesforce reported that its research checkpoint scored 57.21 against EmbeddingGemma 2's 51.07 on a 0–100 scale, a 6.13-point gain. A production-supported checkpoint scored 55.76.

The scores average nDCG@10 across 10 retrieval tasks within the same 41-task MTEB suite. The metric assesses relevance among the top 10 results and gives greater weight to higher-ranked items.

Salesforce evaluated all three models with an 8,192-token context limit. It explicitly labels the measurements as local checkpoint comparisons, not official MTEB leaderboard ranks. The results show improvement over the selected baseline under a controlled protocol, but do not establish a position against the wider field.

A separate autonomous session lasting roughly nine hours targeted serving efficiency. Salesforce reported 36% higher throughput and 17% lower latency with almost no quality loss, without quantifying that change. The company linked those gains to the cost and responsiveness of retrieval across large customer datasets.

Salesforce's engineering record also reports up to 1.84× faster pandas reductions on nullable data types, which explicitly represent missing values. Median hnswlib queries per second rose 4.1% across three index builds, while single-core USearch throughput increased 17.5% with GCC 11.4. Both hnswlib and USearch perform vector search, a central operation in AI retrieval.

The company's nullable-dtype pandas patch covered 134,064 differential cases and recorded zero differences in values, masks, data types or exceptions. Inputs outside its fast path fall back to pandas' existing implementation, and Salesforce submitted all three pandas optimizations as upstream pull requests.

The hnswlib patch reuses scratch memory per worker while preserving byte-identical search results. The USearch patch caches stored vector norms, adding four bytes per vector while preserving recall in the reported evaluation.

Takeaway: The selected tests show measurable gains, but they do not establish the model's standing against the wider field.

What still blocks customer deployment?

Under Salesforce's design, researchers still set the goals, budgets, evaluations and data constraints. Models intended for product use draw only on training datasets that have cleared ethical and legal review, and an improved model must also satisfy production trust and safety requirements.

Salesforce said production work is underway, so SFR-Embedding v4's research results do not amount to a completed customer rollout. The company describes the immediate significance as compression of the controlled research cycle from weeks to days.

Takeaway: The next test is whether the gains pass production review and reach the retrieval systems supporting Salesforce's enterprise agents.