A Lift Is Not a Selection Gradient: What the Amplification-Lift Gap in Thinking Models Measures, Why an Answer-Aware Label and a Pooled Contrast Account for Much of It, and the Within-Prompt Test That Would Show Training Is Misdirected

@misc{mahendrakar2026lift,
  author       = {Mahendrakar, Pranay},
  title        = {{A Lift Is Not a Selection Gradient: What the Amplification-Lift Gap in Thinking Models Measures, Why an Answer-Aware Label and a Pooled Contrast Account for Much of It, and the Within-Prompt Test That Would Show Training Is Misdirected}},
  year         = {2026},
  month        = nov,
  date         = {2026-11-04},
  publisher    = {Life of Research},
  howpublished = {Preprint},
  doi          = {10.5281/zenodo.23951230},
  url          = {https://lifeofresearch.com/preprints/187},
  abstract     = {(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Life of Research Reasoning-oriented post-training makes model traces look more deliberative, and a 2026 study asks whether it amplifies the behaviours that go with correct answers. Annotating 15,282 traces from 15 models, it finds that thinking models express self-correction, hypothesis testing and uncertainty several times more often than instruction-tuned models, while the behaviours with the largest accuracy difference between traces that show them and traces that do not, knowledge alignment and confidence calibration above all, are no more common (self-awareness only modestly so, 1.1 to 1.9 times across six benchmarks). It reads this as training that rewards surface form over the behaviours that predict success. In this analysis the author reviews and synthesises published results to show that the gap, as measured, is not yet evidence that training is misdirected. Three of its ingredients predict much of it without any misdirection. First, the top-ranked labels were assigned by a judge that saw the answer key: the author derives from the study's tables that at least 98 percent of incorrect thinking-model traces that voice doubt are still labelled uncalibrated, so the label behaves like "confident and correct". Second, the headline difference is pooled across questions; within a question, in the study's own control, self-correction is associated with 26.7 points higher accuracy against 30.5 for calibration, which ranks third of nine behind knowledge alignment (59.6) and goal tracking (40.5). Third, outcome-rewarded training acts on within-prompt covariance with reward, and multiplying that control by q(1 - q), where q is prevalence, gives near-equal covariances for self-correction, calibration and goal tracking (0.066, 0.068 and 0.064), against 0.0015 for uncertainty acknowledgment; the largest, 0.099, belongs to knowledge alignment, an outcome-referenced label like calibration. Goal tracking, which is not amplified, is not explained by these three ingredients and counts against this reading. What remains is narrower and real: uncertainty acknowledgment is amplified about five-fold with no within-question association, which imitation of teacher traces, length and clipping biases, hitchhiking on shared tokens, or a training-time role all predict. The paper consolidates 28 published results and two derived quantities in a table and a figure, states the predictions that separate six accounts, gives a reading procedure, and specifies the blinded, within-prompt and stage-separated tests that would settle the claim.},
  keywords     = {reasoning models, chain of thought, reinforcement learning from verifiable rewards, self-correction, confidence calibration, process supervision, selection differential, evaluation methodology},
  language     = {en},
  copyright    = {Creative Commons Attribution 4.0}
}