A new study shows large reasoning models make up their mind almost immediately, then keep talking anyway.
Researchers analyzed the step-by-step reasoning traces that chain-of-thought systems produce before answering. By tracking a model's confidence in its final answer after each reasoning step, they found a sharp commitment boundary where confidence jumps from unstable guessing to a near-final decision, usually in a single step, often long before the reasoning block actually ends. Everything the model writes after that point is what the researchers call epiphenomenal: text that looks like reasoning but doesn't change the eventual answer. Using lightweight probes trained on the model's internal activations, the team detected this transition point with high accuracy, and the method held up even on reasoning tasks the probes hadn't seen before.
That matters because chain-of-thought reasoning is the main lever companies use to make models smarter at inference time, and it's expensive: more reasoning tokens mean more compute and more latency. If a meaningful chunk of that reasoning is decorative, as this paper suggests, the industry's current scaling knob is partly theater. The researchers quantified it, cutting reasoning length by up to 55% by exiting at the commitment boundary, with no meaningful hit to accuracy.
It's a tidy rebuttal to the assumption that longer chains of thought mean more actual thinking rather than more expensive narration.