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

2024-06-06

arXiv.org

AI Summary

Downstream capabilities of frontier AI models—such as multiple-choice question-answering accuracy—exhibit unpredictable scaling behavior, undermining reliable performance forecasting. We identify that this unpredictability stems primarily from non-stationary fluctuations in the probabilities assigned to incorrect options as model scale increases, a phenomenon overlooked by conventional scaling laws that model only the correct-option probability. Method: We introduce, for the first time, the concept of “distractor scaling laws” and develop a multi-step analytical framework linking negative log-likelihood (NLL) to accuracy. This framework enables systematic characterization of co-variation between correct- and incorrect-option probability mass. Contribution/Results: Empirically validated across five model families and twelve multiple-choice benchmarks, our analysis reveals stable NLL scaling—reflecting global probability concentration—while downstream accuracy degrades due to noise amplification inherent in small-set comparative decisions. Our work establishes a new theoretical foundation and modeling paradigm for predictable downstream evaluation.

Technology Category

• Machine Learning: Scalability of ML Systems
• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models
• Reasoning under Uncertainty: Relational Probabilistic Models

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

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 extit{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.

Problem

• Predicting AI downstream scaling remains elusive.
• Degradation in statistical relationship affects predictions.
• Scaling laws for incorrect choices are achievable.

Innovation

• Transforms negative log likelihoods
• Identifies metric degradation mechanism
• Studies probability mass covariation