Inverse Scaling Prize

A two-round public contest (organized by McKenzie, Lyzhov, Pieler, Parrish, Bowman, Perez and collaborators, 2022–2023) that offered prize money for finding tasks on which large-language-model performance gets worse, not better, as model scale increases. The contest produced a fixed set of 11 winning tasks — a reusable dataset (inversescaling.com/data) that recurs across multiple independent papers rather than being tied to one.

Matters to this vault as the shared empirical anchor for a small cluster of task-taxonomy papers: McKenzie et al. (2023) classified the 11 tasks into four causal categories, and Wei, Kim, Tay & Le (2023) re-ran the identical 11 tasks at larger scale and found most of them reverse into U-shaped curves — two papers studying the same dataset reaching a scale-dependent, not fixed, verdict on it.

References

• claim-mckenzie-2023-inverse-scaling-four-causal-categories
• claim-wei-2023-inverse-scaling-becomes-u-shaped-at-larger-scale
• Captures: 10-inbox/raw/2026-10-23-does-published-ml-research-go-further-than-hu.md