视频选题榜 2026-10-02

  • https://arxiv.org/abs/2609.39982 · Mid-Harness · Scaling Actions Between Model and Harness for Terminal Agents · 把 test-time compute 从生成端挪到 model-harness 边界,候选 action 先验后发,强 verifier 让弱 generator 翻盘 50.00% → 68.03% · round=R1