Salesforce unveils SFR-AutoR&D: agents that run the full R&D loop, 1.84x speedups on tuned code
silviocinguetta · x · 2026-09-26
Salesforce AI Research introduced SFR-AutoR&D, specialized agents that autonomously run the research-to-engineering loop: hypothesis formation, controlled experiments, and evidence-based validation.
- Components: AutoR&D-Engineer (code optimization), TrainForge (invents training methods), AutoInfra (supporting infrastructure)
- Results: 1.84× speedup on heavily optimized pandas reductions, +17.5% QPS on USearch, 3.14× faster generation in their RL stack
- Training: +11 to +14pp on held-out math, code, and research benchmarks; full-parameter training of 1T models at 1M-token context
- Every claim is backed by controlled experiments against baselines, discarding what doesn't hold up
More from coding & agent
- Claude Opus 5.5 one-shots a promo video in Cursor: has video animation been solved? — JosephJacks_ · 2026-09-26
- dhh: Opus 5.5 one-shot ported Rust screensaver engine to x86-64 asm, 17x faster — zack_overflow · 2026-09-26
- Anthropic details path to recursive self-improvement as engineers ship 8x more code — dlwh · 2026-09-26
- Making non-black-box embeddings: ask an LLM questions instead of using 1536-dim vectors — kieranklaassen · 2026-09-26
- Font lets you read AI outputs exactly as the AI wrote them, installable in Discord and Slack — amplifiedamp · 2026-09-26
- Only Then Labs Building Provenance-Bounded Context Layer for Agent Memory Sharing — tallmetommy · 2026-09-26