AI models that pause to double-check their own answers are often just going through the motions.
Researchers built a taxonomy of reasoning steps - five groups and seventeen categories - grounded in human cognitive psychology, then used it to study how today's large reasoning models actually think. To do that at scale, they built an automated annotation tool called CAPO and used it to label 277,534 individual reasoning steps, reaching agreement levels close to human expert annotators. The standout finding: the post-answer double-check moments models insert before a final response rarely change that response. When the researchers instead forced models to produce richer, more substantive reflection, failed self-corrections improved markedly. The team says the same pattern shows up in a newer reasoning model and in coding tasks, not just the systems they started with.
This matters because reasoning models are marketed on the promise that they think before they speak, catching their own errors along the way. If that checking step is mostly performative, anyone leaning on self-correction to catch mistakes in code review, math, or research assistance is trusting a safety net that often isn't there. The fix the researchers found - explicitly prompting for deeper reflection - is simple enough that model builders could bake it in today.
The gap between a model that claims to have checked its work and one that actually did tracks with a complaint chain-of-thought researchers have raised for over a year: what a model shows you is not necessarily what it used to reach an answer.