SLDBench v2发布:Claude Opus 5.5仅用4%算力成本超越经典拟合基线

jiqizhixin · x · 2026-11-16

Researchers from CUHK-Shenzhen, HKU, Tsinghua, Peking University and collaborators introduce SLDBench v2 (arXiv preprint) to benchmark active scaling-law discovery: agents see a catalog of experiments with prices, choose which results to buy, refine candidate laws, and decide when to stop. The environment replays recorded training runs at original compute cost and validates predictions on hidden larger-scale experiments. Claude Opus 5.5 reaches median R² 0.570 using just 4.01% of the pool's historical compute, beating the classic reference law fitted on all data (0.455). Across 30 tasks, all eight evaluated systems degrade when forced to actively select experiments versus receiving all outcomes — cheap observations can leave key assumptions untested.

原文链接 →

「模型」频道最新

更多「模型」频道 AI 资讯 →