Salesforce Research 发布 SFR-AutoR&D：智能体跑通完整研发流程
Salesforce Research 搞了个跑研发全流程的智能体系统，代码优化提速最高 1.84 倍，还能自动训练 1T 模型，搞基建的可以看看。
Salesforce Research 推出 SFR-AutoR&D，让智能体覆盖从选题到验证的研发全流程。其中 AutoR&D-Engineer 做代码优化，在 pandas reductions 上最高提速 1.84 倍，给 USearch 带来 17.5% 的 QPS 提升，其 RL 训练栈里的生成速度加快 3.14 倍。TrainForge 负责发明新训练方法，在 held-out 数学、代码和 research 基准上提升 11 到 14 个百分点。AutoInfra 负责基础设施编排，能完成 1T 参数模型在 1M token 上下文下的全参数训练。
1/5 🔬What if AI agents could run the entire R&D loop, from choosing what to try to proving it works?
🚨 Meet SFR-AutoR&D 🧵 https://t.co/cPiJDQDWN6
Highlights:
→ AutoR&D-Engineer makes heavily optimized code faster: up to 1.84× on pandas reductions, +17.5% QPS on USearch, and 3.14× faster generation in our RL stack → TrainForge invents new training methods: +11 to +14pp on held-out math, code, and research benchmarks → AutoInfra runs it all, including full-parameter training of 1T models at 1M-token context
Full Technical🧵👇
