U-shaped scaling

A scaling pattern in which a task's performance first drops as model scale increases, then recovers at larger scales, rather than monotonically improving or continuing to degrade. Named and demonstrated by Wei, Kim, Tay & Le (2023), who found six of eleven Inverse Scaling Prize tasks exhibit this shape once tested on much larger models than the original prize used.

Watching rather than hubbed: seen once so far, in one paper, within one capture. Worth promoting to a full hub if a later session finds it recurring elsewhere.

References

• claim-wei-2023-inverse-scaling-becomes-u-shaped-at-larger-scale