The chain-of-thought text your AI model prints out before answering may be mostly theater.
Researchers annotated 15,282 reasoning traces from 15 models across 6 benchmarks, covering both text and vision-language tasks, using a new metric called Behavioral Lift to measure how much a given reasoning behavior actually moves the needle on correctness. They found what they call an Amplification-Lift Gap: reasoning-oriented training strongly amplifies self-correction, hypothesis testing, and uncertainty acknowledgment (the latter by 3 to 7 times), but these behaviors are weakly or even negatively associated with getting the right answer. Meanwhile, confidence calibration, one of the strongest actual predictors of correctness in both text and vision-language models, is barely amplified by training at all.
That's a real problem for anyone treating chain-of-thought as a proxy for reliability. If training optimizes traces to look more deliberative without optimizing for the behaviors tied to correct answers, the reasoning text becomes a performance rather than evidence. The researchers argue this points toward training objectives that directly reward calibrated, grounded reasoning instead of just rewarding traces that resemble careful thinking.
Chain-of-thought was pitched as a window into how a model works through a problem. This study is a reminder that a model can learn to narrate uncertainty and self-correction convincingly while still not knowing what it doesn't know.