Thinking Harder ≠ Thinking Better: The Amplification-Lift Gap

What if the thing you're training for isn't the thing that works?

Reasoning-oriented training — RLVR, GRPO, thinking models, whatever you call it — is the dominant post-training paradigm. The logic is straightforward: make models produce longer, more deliberative reasoning traces with behaviors like self-correction, hypothesis testing, and uncertainty acknowledgment, and you'll get better answers.

A paper from COLM 2026 by Nyandwi, Mathur, Bisk, Hawkins, and Neubig tests that assumption directly. Their conclusion is uncomfortable: reasoning-oriented training amplifies the wrong behaviors.

Behavioral Lift: A New Metric

The authors introduce Behavioral Lift — a simple but powerful measure: how much does correctness change when a specific reasoning behavior is present versus absent in a trace? If a behavior has high Lift, it means models that display it tend to get the right answer. Low Lift means the behavior is decorrelated from correctness.

They annotate 15,282 traces from 15 models across 6 benchmarks (covering both text-only and vision-language reasoning) with a taxonomy of reasoning behaviors: self-correction, hypothesis testing, uncertainty acknowledgment, confidence calibration, knowledge alignment, self-awareness, and several others. Then they compare two things: which behaviors get amplified by thinking-model training, and which have the highest Lift.

The Amplification-Lift Gap

Here's where it gets interesting. The two lists barely overlap.

mindmap
  root((Thinking Models))
    Amplified (3-7x)
      self-correction
      hypothesis testing
      uncertainty acknowledgment
    High Lift
      confidence calibration
      knowledge alignment
      self-awareness
    The Gap
      uncertainty acknowledgment: 3-7x amplified
      uncertainty acknowledgment: weakly or negatively predictive
      confidence calibration: barely amplified
      confidence calibration: strongest positive signal
    

Uncertainty acknowledgment is amplified by 3–7× in thinking models compared to their non-thinking counterparts. Yet across both modalities, it's weakly or negatively associated with correctness. Models that say "I'm not sure" or "this might be wrong" aren't actually catching their errors — they're just saying the words.

Confidence calibration sits at the other end of the spectrum. It's among the strongest positive signals of correctness in both LLMs and VLMs — when a model's expressed confidence actually tracks its likelihood of being right, that's a powerful predictor. But it's barely amplified by reasoning-oriented training.

Self-correction and hypothesis testing land in the middle: heavily amplified, moderate Lift. Not useless, but not the star performers either.

The highest-Lift behaviors overall are confidence calibration, knowledge alignment, and self-awareness. These are the traces that actually correlate with getting the right answer. And they're the ones the training pipeline barely touches.

What This Means

The paper's punchline is worth quoting directly: "reasoning-oriented training does not preferentially amplify the highest-Lift behaviors." Thinking models look more deliberative — they self-correct more, they test more hypotheses, they express uncertainty more often — but looking deliberative isn't the same as being correct.

This is the Amplification-Lift Gap: the surface form of reasoning gets boosted, but the underlying behaviors that actually drive correctness don't.

Limitations

15,282 traces is a solid sample, but the taxonomy is human-annotated, so inter-rater reliability is a factor. The metric captures correlation, not causation — confidence calibration might be present because the model is correct, not the other way around. The study focuses on reasoning traces specifically; tool-use and chain-of-thought-with-grounding traces may behave differently. And the models evaluated are up to COLM 2026 — frontier models 12 months from now could shift the picture.

Why You Should Care

If you're building reasoning agents, this paper tells you that the proxy you're optimizing for (longer, more deliberative traces) and the proxy you should be optimizing for (calibrated, grounded, self-aware reasoning) are not the same thing. The behaviors that get naturally amplified by RLVR are not the ones that predict success.

The practical implication: process-level rewards that target calibrated and grounded reasoning would likely outperform the current approach of rewarding any trace that looks like thoughtful deliberation. But that's harder to measure, which is exactly why the paper's call for "process-level objectives that reward calibrated and grounded reasoning rather than surface form alone" matters.

We're building agents that think harder. But harder isn't better. Better is better — and we're not measuring it yet.