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

Predictable behavior from scaling advanced AI systems is an extremely desirable property for engineers, companies, economists and governments 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 shines a light on 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 show 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 reveal 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 specific 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.

Predicting how advanced AI systems will perform on specific tasks like multiple-choice question answering as they get bigger (scale) is surprisingly difficult, even though we can predict their general training progress well. This paper reveals a key reason: current scoring methods for these tests obscure the true relationship between AI scale and performance. The core issue is that these tests require the AI to not only identify the correct answer but also to distinguish it from a few specific incorrect options. To accurately predict performance, we need to understand how the AI's confidence in both the right and specific wrong answers changes as it scales. The authors show that by analyzing these dynamics, we might be able to develop predictable "scaling laws" even for these tricky incorrect choices. This research helps explain why predicting downstream task performance is harder and offers a path towards making evaluations of new AI models more reliable.