AI
- arXiv:2610.30113 — Quant-Arena — New benchmark testing whether LLM agents can run end-to-end quantitative trading research on real A-share data; none of 9 evaluated models beat the equal-weight market in the sealed 2024 test year.
- Hugging Face Papers — Automating ML for Science — Frontier coding agents can beat published best results on scientific ML benchmarks, but the gains come mainly from routine engineering tricks rather than domain insight, and lightweight planning/experiment-management skills substantially boost performance.
- Hugging Face Papers — Harness Scaling for Generalizable Agentic RL — Training an agentic model across structured variations of its execution harness, rather than one fixed harness, improves generalization to unseen harness configurations, including external black-box harnesses.