Amplified Does Not Mean Predictive: Reasoning Behaviors in Thinking Models

Title: Amplified Does Not Mean Predictive: Reasoning Behaviors in Thinking Models Source: arxiv Measuring Behavioral Lift across 15 models and 6 benchmarks reveals an Amplification-Lift Gap in reasoning-oriented training. While thinking models increase self-correction and hypothesis testing, these amplified behaviors correlate less with correctness than high-lift behaviors like confidence calibration and knowledge alignment.