THE APEX TIMES
The company says its new autonomous research system can run the “scientific loop” for code and embedding models, shrinking upgrade cycles from weeks of human experimentation to about three days, while keeping improvements only when they clear evidence and safety constraints.
Salesforce is bringing more autonomy to how enterprise AI models improve, unveiling SFR-AutoR&D, a set of research agents designed to run the work of hypothesis, experimentation, and model upgrade in a loop. The company said the system can train and refine embedding models used in production workflows, and that it retains changes to code and model training only when they outperform defined baselines and meet accuracy and safety requirements.
At the center of the effort is the research loop itself. Salesforce says SFR-AutoR&D takes a goal, a target system to improve, and a compute budget, then formulates hypotheses, implements methods, and runs controlled experiments against baselines. The system can hold itself to evidence by checking whether improvements are supported by results, rather than simply testing changes that look promising on the surface.
Salesforce said the agents have already been used to upgrade its production AI models. In one example, the company says the system autonomously improved SFR-Embedding V3, the embedding model behind retrieval in Data Cloud, and then moved to create SFR-Embedding V4. In another demonstration, Salesforce pointed the agents at the open-source embedding model EmbeddingGemma 2 and reported measurable improvements in both model quality and latency, again culminating in a better embedding model.
The company also described results from prior work the system has done on public codebases that many data teams rely on. Salesforce said the first published results involved pandas, the widely used Python library for manipulating large data tables, along with hnswlib and USearch, libraries used for similarity search that supports retrieval in AI applications. Salesforce characterized these as heavily optimized projects and said the agents still found additional gains, each supported by benchmark evidence.
Embedding models are the unglamorous workhorses of enterprise AI, Salesforce said. They convert text into numerical vectors, allowing AI systems to find relevant records, documents, or case histories out of large pools of customer data. When an AI agent answers a question grounded in a customer’s own information, the retrieval step depends on embedding quality, because the embedding model determines what information comes back for the agent to use next.
Salesforce contrasted the timeline of traditional model improvement with what SFR-AutoR&D can do. Historically, the company said, improving embedding models has largely been a human process, with scientists proposing an approach, running experiments, reading results, and iterating over weeks. Salesforce said its agents complete the cycle in about 80 hours, just over three days, framing the change as an acceleration of scientific experimentation rather than a removal of scientific discipline.
To support autonomous improvement, Salesforce said it built two agents with different roles. AutoR&D-Engineer searches for algorithmic and systems improvements inside existing code, while TrainForge explores new training methods and evaluates candidates against held-out benchmarks that the training process does not see. Both agents operate on AutoInfra, the training backbone Salesforce says carries experiments across shared GPU capacity.
In evaluation, Salesforce reported specific performance and efficiency outcomes. The company said SFR-AutoR&D was assessed against 10 public retrieval tasks in MTEB and improved the model’s ability to place relevant results near the top, raising its nDCG@10 score by 6.13 points. It also reported a separate session focused on speed and cost, where Salesforce said the system delivered 36% more throughput and 17% lower latency with almost no loss in quality.
Salesforce said trust is built into the problem definition rather than bolted on afterward. The company said models trained for product use draw only on datasets that have cleared ethical and legal review, and that even if a model performs well in evaluation, it must still clear production requirements before it reaches customers. The company also emphasized that autonomy has boundaries: researchers set the goal, budget, evaluation, and data constraints, and the system does not retain improvements unless they are verified against baselines.
Looking ahead, Salesforce said self-improving systems are likely to become normal in enterprise AI development, with researchers shifting toward setting constraints and goals for autonomous frameworks. The company also referenced a related effort called CodeEvolve, which uses evolutionary search with large language models to improve performance and code quality in Java and Salesforce Apex, and indicated an upcoming blog would cover that work in more detail. For customers and the market, the key question will be how quickly these lab results translate into broader, ongoing improvements in real customer-facing workloads, and what operational guardrails Salesforce continues to refine as autonomy expands.
A strong quarter from CoreWeave, a major cloud provider for AI compute, helped quiet worries that technology infrastructure spending could cool. Shares tied to AI chips and networking moved higher alongside the sentiment shift.
A major downgrade cut Nvidia’s target price by about 60 percent, while other Wall Street analysts held bullish outlooks and pointed to higher upside targets.
A Wall Street note captured a split view of Advanced Micro Devices, with higher profit expectations offset by a lower price target amid recent market pressure.
A Yahoo Finance report says Cathie Wood’s ARK Investment Management reduced its Amazon position while adding to a basket of aerospace-related names after strong trading activity around a rocket launch company.
The new Gemini Decisions API is designed to score bounded yes/no, multiple-choice, and rubric questions across text, images, audio, video, and PDFs in a single pass, with confidence scores and pricing aimed at low-latency decisioning.
The browser-based retro video game debuts on Googlebook, with Google pledging $11 per play to Movember men’s health causes through November 30, up to $500,000.
An equity note circulated by Yahoo Finance points to AI chip testing and a widening Nvidia relationship as key catalysts for Teradyne, the maker of semiconductor test equipment.
The comments, reported by Yahoo Finance, put a spotlight on how cloud AI infrastructure providers are weighing Nvidia’s role as scrutiny continues around the broader semiconductor push for artificial intelligence.
In a portfolio-related update cited by Yahoo Finance, the veteran investor said he was positioning for a near-term lift in Nvidia after the stock’s recent selloff, while also questioning the durability of the company’s competitive advantage.
A market report highlights how Anthropic’s momentum in generative AI could reshape competitive dynamics for Microsoft’s cloud and software stack, adding pressure to an already crowded field.
Comments
Community guidelines. Comments may be removed for violations.
Loading comments...