When Thinking Longer Makes You Wronger: Inference Sampling Amplifies the Learned Ordering

Extra inference is often described as giving a model more opportunity to correct itself. But repeated sampling cannot, by itself, supply a distinction the active model has not learned. It estimates that model's ordering of candidate outcomes more faithfully. We test the resulting prediction: the value of added inference sampling should depend on whether the ordering being amplified is aligned with the environment. The principal experiment uses four controlled decision families, three neural-model widths, budgets from 1 to 64 samples per action, 16 untouched confirmation seeds, and 76,800 unique held-out item–seed cases. Models receive native variables only. On the same held-out items, a narrow-history model's ranking is opposed to the world, while relevant learning changes the specialist's ranking into alignment. Increasing sampling then reverses sign: it raises light-model regret and lowers specialist regret in all four families and in all 64 seed–family cells. The paired reversal ranges from 0.2848 to 0.2900 expected-regret units. A parameter-, data- and optimisation-matched learner trained on shuffled outcomes does not rescue the same items; its deficit is largest, 0.7377, in the family built around a reversing spurious feature. Across learning-history coverage, ranking accuracy rises from 0.0846 to 0.8503 and the compute effect moves from +0.1564 (harm) to −0.1013 (benefit). The curve crosses slightly before oracle-action mastery: with three actions and unequal regrets, the full ordering—not top-action accuracy alone—sets the finite-budget effect. An exact two-action analysis supplies the limiting case. Supporting experiments show that the same narrow model benefits in its ordinary environment, that increased width does not replace the missing relation, and that a globally better autoregressive specialist need not improve under plurality voting when its itemwise rankings remain mixed. The result is therefore not that more compute is generally harmful. Inference sampling is an amplifier of the active model's ordering. Whether that amplification helps depends on learned structure and the regret geometry of the decision.