Salesforce SFR-AutoR&D delegates the model research loop to agents
Salesforce AI Research says SFR-AutoR&D improves embedding models by assigning hypothesis generation, implementation, controlled experiments and validation to research agents.
AT A GLANCE
Key points
- SFR-AutoR&D agents generate, implement and validate model research experiments.
- Salesforce reports a 6.13-point average nDCG@10 gain across 10 retrieval tasks.
- A separate session reported 36% more throughput and 17% lower latency.
What is changing?
Salesforce AI Research has introduced SFR-AutoR&D, a system that manages the model-research loop from hypothesis to validated improvement. AutoR&D-Engineer agents generate and implement experiment ideas, retaining only changes supported by measurements, while TrainForge runs training and evaluation jobs on AutoInfra. Salesforce says it used the approach on work spanning SFR-Embedding V3 and an EmbeddingGemma 2 base model to SFR-Embedding v4.
According to the company, an approximately 80-hour research cycle improved average nDCG@10 by 6.13 points across 10 MTEB retrieval tasks. A separate nine-hour session delivered 36% higher throughput and 17% lower latency with almost no quality loss. These figures come from Salesforce's own evaluations; people still define the objective, data, budget and production constraints.
Explore the original sources
Source date: 11 November 2026.
Published by AI Radar: 12 Nov 2026, 00:16 (Istanbul, UTC+3).
FOLLOW THE TOPIC
Topics and tags
Share your discovery.
Send this story to a friend.
Reader comments
What do you think about this development? Share your experience or question. Comments are published after editorial approval.
Your name/nickname and comment become public if approved. Read: Privacy notice.
Loading comments…