1
Why external scaffolds aren't enough. Chain-of-thought, self-consistency and search improve the outer procedure, but never check whether the model's internal pathway matches the logic each step needs.
A guided tour of LoTR
The problem, the idea, the method, and the results — an animated walkthrough of Logic-of-Thought Routing, built from the paper.
What the tutorial covers
2
How LoTR routes logic. A compact basis of recurring logic states, per-layer probes that read the current state, and cached head-templates that are softly blended into a gate matrix — online.
3
That it works and stays cheap. +2.66 points (+5.37% relative) on the Llama 8×8 study — helping or tying 51 of 64 cells — with only +0.15% tokens, and positive transfer to Qwen2.5-14B and Mixtral-8×7B.