Why Has Predicting Downstream Capabilities of Frontier AI Models with Scale Remained Elusive?

TL;DR

This paper investigates why predicting the scaling behavior of downstream capabilities in advanced AI models remains difficult, identifying key factors affecting predictability and proposing directions for more reliable evaluation methods.

Contribution

It reveals how downstream performance metrics degrade the statistical relationship with scale and suggests that scaling laws for incorrect choices could improve predictability.

Findings

Downstream performance metrics involve complex probability fluctuations.

Scaling laws for incorrect choices may be more predictable.

Downstream evaluation predictability can be improved by understanding probability mass fluctuations.

Abstract

Predicting changes from scaling advanced AI systems is a desirable property for engineers, economists, governments and industry alike, and, while a well-established literature exists on how pretraining performance scales, predictable scaling behavior on downstream capabilities remains elusive. While many factors are certainly responsible, this paper identifies a significant factor that makes predicting scaling behavior on widely used multiple-choice question answering benchmarks challenging and illuminates a path towards making such downstream evaluations predictable with scale. Using five model families and twelve well-established multiple-choice benchmarks, we demonstrate that downstream performance is computed from negative log likelihoods via a sequence of transformations that progressively degrades the statistical relationship between performance and scale. We then pinpoint the mechanism causing this degradation: downstream metrics require comparing the correct choice against a small number of specific incorrect choices, meaning accurately predicting downstream capabilities requires predicting not just how probability mass concentrates on the correct choice with scale, but also how probability mass fluctuates on the alternative incorrect choices with scale. We empirically study how probability mass on the correct choice co-varies with probability mass on incorrect choices with increasing compute, suggesting that scaling laws for \textit{incorrect} choices might be achievable. Our work also explains why pretraining scaling laws are commonly regarded as more predictable than downstream capabilities and contributes towards establishing scaling-predictable evaluations of frontier AI models.