Research notes from a project on scalable oversight: training an AI to break hard tasks into subtasks that weaker models, or humans, can reliably judge.https://mai-alignment.github.io/superalignment-task-decomposition/https://mai-alignment.github.io/superalignment-task-decomposition/updates/1_baselines/https://mai-alignment.github.io/superalignment-task-decomposition/updates/1_baselines/The system in which reasoning LLM decomposes complex tasks into simpler ones for a weaker LLM to implement loses with single zero-shot thinking LLM but wins with zero-shot non-thinking model. This leaves a lot of room for improvement.Wed, 04 Nov 2026 00:00:00 GMThttps://mai-alignment.github.io/superalignment-task-decomposition/updates/0_vision/https://mai-alignment.github.io/superalignment-task-decomposition/updates/0_vision/Our vision and goals for scalable oversight over AI systems.Tue, 03 Nov 2026 00:00:00 GMT