The Termination Circuit (how reasoning models stop thinking).

A LessWrong post details research into why reasoning models like OpenAI's o1 overthink their answers. Researcher Chandram Dutta tested Qwen3-1.7B solving math problems and found something striking: the model reaches the correct answer about halfway through its reasoning chain, then keeps going anyway. Dutta traced the stopping decision to a small cluster of neural layers he calls the termination circuit. Rather than measuring whether the reasoning is sound, this circuit verifies a simple condition: does the model's written answer match what it computed internally? Here's the puzzle: the circuit can't be easily steered with a simple signal. The stopping trigger is encoded in a high-dimensional pattern, not a single dial you can turn. For building more efficient reasoning models, Dutta's suggestion is direct: the real lever isn't improving the circuit itself, but getting the model to state its answer much earlier, letting verification happen sooner.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton