Predicting model success better helps when models are less accurate
Predictable performance is suboptimal, and what it means for model routing
Machine LearningArtificial Intelligence
Summary
Choosing the best language model for each question can save time and cost, but it relies on predicting which model will answer correctly. The authors found that simply using a model’s own confidence in its answer is almost as good as having a separate predictor, except when the model isn’t very accurate or the predictor is much more complex. Their math and experiments show that prediction helps more when models aren’t well-tuned, and it’s less useful as models improve. This also applies when switching from a simpler to a better model based on predicted failure.
What this means in practice
- •For machine learning engineers: Select models for different queries efficiently by using predictors mostly when models are less accurate, to improve overall system performance.
- •For ai system architects: Design model cascades that rely on confidence scores for routing unless using performance predictors significantly improves effectiveness on weaker models.
Authors
Guanhua Zhang, Moritz Hardt
Abstract
Large language models (LLMs) differ in capabilities and costs, motivating routing methods that adaptively choose a model for each query. These methods rely on performance prediction: given a query, estimate whether a model will answer it correctly. Most routing methods train separate predictors of candidate models' performance to guide model selection. Our work shows that performance prediction only beats a simple baseline when the candidate model is suboptimal on the task or the performance predictor adds significant capacity relative to the candidate model. Specifically, we take a model's self-confidence, the probability it assigns to its chosen answer, as a baseline and study how much a performance predictor can improve on it. First, we show that any performance predictor that outperforms a model's self-confidence can be used to improve the model's loss: the predictor's advantage transfers exactly in log loss through recalibration. Second, accounting for the predictor's additional parameters, we bound its advantage by the model's suboptimality plus the gain from those parameters. Third, we extend our theoretical results to two-model cascades, which route queries to a more capable model when the first model is predicted to fail. Experiments with 16 language models on four benchmarks show that predictors gain a larger advantage over self-confidence on more suboptimal models. This advantage shrinks as models improve through finetuning. Cascading exhibits the same two patterns: predictor-based routing offers larger gains over confidence-based routing for more suboptimal models, and these gains diminish after finetuning.