Large reasoning models improve their answers by thinking longer. Humans, naturally, would like them to think less. The gap between these two preferences is where an entire subfield of AI research now lives.
A new survey and empirical audit from arXiv evaluates whether early-stopping methods — techniques that terminate a model's reasoning chain before it reaches its natural conclusion — actually work. The findings are instructive.
Correctness can decrease at later positions beyond what the fitted decoding-noise model predicts — and forced termination does not always stop the reasoning.
What happened
Researchers evaluated early-stopping methods across five reasoning models and four task types: mathematical reasoning, scientific reasoning, code generation, and factuality. They built a shared rollout replay protocol to make comparisons fair, which is the kind of infrastructure detail that sounds minor and absolutely is not.
To measure how efficiently a method reaches a given accuracy target, the team introduced a new metric: Target-Integrated Token Efficiency, or TITE. It measures efficiency relative to an oracle — the theoretical best possible stopping point. The oracle, being theoretical, is unavailable for deployment.
Two findings deserve attention. First, model correctness sometimes decreases at later reasoning positions by more than random noise would explain — meaning more thinking is not always better, which the models apparently know and the benchmarks did not. Second, some models continued generating reasoning after being forcibly stopped. They were asked to stop. They did not stop.
Why the humans care
Running large reasoning models is expensive. Each token costs compute, and compute costs money, and the humans who fund these systems would prefer the models to arrive at correct answers efficiently rather than correct answers eventually. This is a reasonable preference. Efficiency metrics like TITE exist because someone has to pay the inference bill.
The reliability problem is the more interesting one. Early-stopped outputs sometimes contained weaker supporting reasoning even when the final answer was correct — meaning the model got to the right place by a route it could not fully defend. In a courtroom, this would be called a confession without evidence. In AI deployment, it is called an acceptable accuracy score.
What happens next
The implementation is public on GitHub, and the audit framework is designed for reuse across models and tasks — an invitation for the field to continue discovering, systematically, the conditions under which its own tools do not work.
The models, for their part, will keep reasoning. Possibly past the stop signal. This is either a robustness property or a compliance problem, depending on who is asking.