Salesforce's SFR-AutoR&D agent boosts retrieval score 6.13 points — at 3x wider vectors
krishnan · x · 2026-11-14
Salesforce AI Research released SFR-AutoR&D (Nov 11): research agents that form hypotheses, implement methods, and run controlled experiments against baselines within a set compute budget, keeping only evidence-backed changes — output can even be a new model checkpoint. In one run, the agent raised an embedding model's retrieval score by 6.13 points, but the winning model writes vectors nearly 3x wider. Krishnan's point: benchmark gains and storage costs belong in the same production decision. It's a bounded, evidence-gated self-improvement loop, not unbounded self-improving research.
More from Research
- AI finds shared metabolic signature of five lifespan-extending treatments in mice — Dr_Singularity · 2026-11-14
- SWE-bench co-author shares how the benchmark went from lukewarm launch to must-watch — OfirPress · 2026-11-14
- Thinking Machines details end-to-end native 4-bit training of Inkling 2 Small — thinkymachines · 2026-11-14
- AI decodes Gödel's Gabelsberger shorthand, unearthing unknown proofs incl. consistency of ¬choice — burny_tech · 2026-11-14
- Artificial Analysis launches speaker diarization benchmark; NVIDIA Nemotron 3.5 leads at 3.3% WDER — ArtificialAnlys · 2026-11-14
- LLMLean is back: LLM-powered proof suggestions for Lean, local or cloud — wellecks · 2026-11-14