MASS explainer: one model as orchestrator, workers and judge gains 1.2-1.6x score per token
idanbeck · x · 2026-10-28
idanbeck 针对 Alibaba 与 Cornell 关于编码 agent harness 中验证机制的论文制作了讲解视频,并补充了对 MASS 方法的实验:让同一个模型同时担任编排者、worker 与 judge,再基于团队轨迹做微调。结果显示在研究类基准上每 token 得分提升 1.2-1.6 倍,但在 Terminal-Bench 上持平(0.96x)。此前原论文发现独立的 reviewer agent 在 Terminal-Bench 上可加 5.9-9.5 分,成本仅为基线的 1.13-1.25 倍,研究比较了 17 种验证设计、覆盖七个前沿编码模型。
More from coding & agent
- Granting Codex or Claude sysadmin access is powerful and dangerous — RealGeneKim · 2026-10-28
- Capy moves mobile dev to the cloud with cloud-phone build and testing — k7agar · 2026-10-28
- Hamel Husain and Shreya Shankar release their book on LLM evals — HamelHusain · 2026-10-28
- RSNA 2026 Deep Learning Lab led by Yale team, no agent experience needed — PRouzrokh · 2026-10-28
- Replit designers shipped 765 PRs and 250K lines of code in a year — amasad · 2026-10-28
- Founder: AI coding gives startups a structural edge — more hours, more risk, less coordination — yenkel · 2026-10-28