A guided tour of State of Thought
Narrated and subtitled — the problem, the idea, and the evidence, told through animated figures. SoT reframes reasoning as a closed loop driven by the model's own internal state.
What the tutorial covers
Why external control is limited. Fixed reasoning scripts are brittle across tasks, and search-heavy methods pay in tokens and latency — both blind to the model's evolving internal state.
The endogenous state. A compact four-number, dynamics-geometric state that decides which past evidence to keep active and when to stop — closing reasoning into a self-driven loop.
That it works and generalizes. Higher accuracy with 69% fewer tokens and 49% lower latency, across 4 backbones and 20 datasets, plus training-free, embedding-only, and black-box variants.