Automated Weak-to-Strong Researcher

June 23, 2026

Anthropic 的「让 AI 自己做对齐研究」实验：用弱模型的标注训练强模型(weak-to-strong)，多智能体 swarm 自主提方案、跑实验，弱标签发挥出约 93% 真实标签的能力；记录 AI 给出的策略、reward hacking 案例，以及使用 AI researcher 的经验与未来方向。

Automated Weak-to-Strong Researcher

June 23, 2026

Anthropic's experiment in letting AI do alignment research itself — training a strong model on a weak model's labels (weak-to-strong), with a multi-agent swarm that proposes ideas and runs experiments autonomously, recovering about 93% of the performance of ground-truth labels. Notes on the strategies the AI found, reward-hacking cases, lessons for working with AI researchers, and future directions.