Carnegie Mellon researchers have identified a fundamental mismatch in how AI reasoning models are trained, finding that the behaviors most amplified during training aren't those most predictive of correct answers.

The study, published in COLM 2026, analyzed 15,282 reasoning traces across 15 models and 6 benchmarks spanning both text-only and vision-language tasks. Researchers introduced "Behavioral Lift," a metric measuring how much correctness changes when specific reasoning behaviors are present versus absent.

The Amplification-Lift Gap

The team discovered what they term an "Amplification-Lift Gap" — thinking models strongly amplify self-correction, hypothesis testing, and uncertainty acknowledgment by 3-7 times, yet these behaviors show weak or negative correlation with correctness.

Meanwhile, the highest-lift behaviors — confidence calibration, knowledge alignment, and self-awareness — receive minimal amplification during reasoning-oriented training. Confidence calibration emerged as among the strongest positive signals of correctness in both text and vision tasks, yet was barely enhanced by current training methods.

The research team, led by Jean de Dieu Nyandwi alongside Leena Mathur, Yonatan Bisk, Robert Hawkins, and Graham Neubig, developed a comprehensive taxonomy of reasoning behaviors applicable to both large language models and vision-language models.

"Reasoning-oriented training can make traces look more deliberative without amplifying the behaviors most tied to model correctness," the researchers noted.

The findings suggest current training approaches prioritize surface-level reasoning appearance over substantive correctness indicators. Models learn to exhibit behaviors that appear thoughtful — like acknowledging uncertainty or testing multiple hypotheses — but these don't translate to better performance.

Implications for AI Development

The research challenges assumptions about reasoning model training, particularly for companies like OpenAI and Anthropic developing advanced reasoning systems.

The study advocates for process-level training objectives that reward calibrated and grounded reasoning rather than surface form alone. This could reshape how frontier AI labs approach reasoning model development, emphasizing behaviors that genuinely improve accuracy over those that merely appear sophisticated.

The paper will be presented at COLM 2026, adding to growing research on alignment between AI training methods and desired outcomes.