McKenzie et al. (2023) classify inverse-scaling tasks into four causal categories — Strong Prior, Unwanted Imitation, Distractor Task, and Spurious Few-Shot

inverse-scalingtask-taxonomyneural-scaling-lawsllm-evaluationinverse-scaling-prize

McKenzie, Lyzhov, Pieler, Parrish et al., "Inverse Scaling: When Bigger Isn't Better" (arXiv:2306.09479; Transactions on Machine Learning Research, 10/2023), reports the results of the two-round Inverse Scaling Prize, a public contest that collected 11 tasks on which large-language-model performance gets worse with scale. The paper goes past merely cataloguing those tasks: "we identify four potential causes of inverse scaling: (i) preference to repeat memorized sequences over following in-context instructions, (ii) imitation of undesirable patterns in the training data, (iii) tasks containing an easy distractor task which LMs could focus on, rather than the harder real task, and (iv) correct but misleading few-shot demonstrations of the task." It labels these "Strong Prior," "Unwanted Imitation," "Distractor Task," and "Spurious Few-Shot," and assigns each of the 11 prize tasks to one category — e.g. Memo Trap and Prompt Injection under Strong Prior, Modus Tollens under Unwanted Imitation, Pattern Match Suppression and NeQA under Distractor Task, Hindsight Neglect and Repetitive Algebra under Spurious Few-Shot.

This is a mechanism-level taxonomy of task-level scale divergence — "this task plateaus/degrades because X," not merely "this task plateaus" — and it predates Hu et al.'s 2026 "Neural Neural Scaling Laws" by three years; Hu et al. cite this paper for the phenomenon itself without extending its causal classification. Wei, Kim, Tay & Le (2023) later re-ran the same 11 tasks at larger scale and found the inverse-scaling trend itself is not stable across model size — only four of the 11 tasks remained inverse scaling, with six reversing into U-shaped scaling. (Wei et al. re-tested the scaling trends, not McKenzie et al.'s causal category assignments.)