Reasoning Models Are Getting Better at Looking Thoughtful—But Not at Actually Being Right

When language models spend more "thinking time" working through problems, their traces look increasingly deliberative. They second-guess themselves, test hypotheses, acknowledge uncertainty. But a critical question emerges: does all this amplified reasoning actually predict correctness? The authors find a surprising disconnect—models are trained to look more thoughtful without learning to reason better.

The Amplification-Lift Gap

The core insight hinges on a new metric called Behavioral Lift, which measures how much a model's correctness improves when a specific reasoning behavior is present in its trace versus absent. This is the inverse of asking whether training amplifies a behavior; a behavior can become more common without being predictive of right answers.

Across 15 models and 6 benchmarks (including both text and vision-language tasks), the authors manually annotated 15,282 reasoning traces using a taxonomy of core behaviors—things like self-correction, hypothesis testing, confidence calibration, and knowledge alignment. The results reveal a systematic mismatch.

High-amplified, low-predictive behaviors:

• Uncertainty acknowledgment is amplified 3–7 times by reasoning-oriented training but weakly or negatively associated with correctness.
• Self-correction and hypothesis testing are strongly amplified yet show modest lift.

High-lift, low-amplified behaviors:

• Confidence calibration (knowing when you're right and when you're wrong) is among the strongest signals of correctness in both modalities yet barely amplified by training.
• Knowledge alignment and self-awareness show similar patterns: genuinely predictive but underamplified.

The asymmetry is stark. Standard reasoning training optimizes for surface-level deliberation—more words, more steps, more hedging—rather than behaviors that actually ground correctness.

Why This Matters

Thinking models (systems like o1) have shown impressive benchmark gains by adding compute-intensive reasoning phases. The natural assumption is that this training directly improves reasoning quality. The authors show that assumption is incomplete. Reasoning-oriented training amplifies certain forms of expression without necessarily teaching models to reason more accurately.

This finding has practical consequences. If a model learns to say "I'm uncertain" without genuinely improving calibration, or to second-guess itself without fixing errors, then scaling reasoning compute may plateau in effectiveness. Surface-form improvements could mask stalled progress on true reasoning.

The authors argue for shifting objectives from rewarding trace length or lexical diversity toward process-level rewards that specifically target calibrated and grounded reasoning—pushing models to amplify behaviors that actually correlate with correctness.

The work is descriptive and diagnostic: the paper identifies the gap but stops short of proposing a complete retraining strategy. Still, it provides a framework (Behavioral Lift) and evidence base (15,282 annotations across modalities and benchmarks) that future work can build on. For researchers tuning reasoning models, the takeaway is clear: measure what matters, not what's easy to measure.