When in-context learners should broaden their model choices
When Should an In-Context Learner Expand Its Hypothesis Space?
Machine Learning
Summary
Learning systems can quickly adjust when working within familiar models, but the challenge is deciding when to try more complex models. The paper looks at how a learner decides whether to stick with simple fixes or expand to a richer model family, balancing the costs and benefits in a step-by-step way. The authors use a formal environment to predict how best to make these decisions and test transformers and language models to see if they follow this ideal approach.
What this means in practice
- •For machine learning engineers: Improve adaptive AI systems by deciding when to switch to more complex models after failure, optimizing prediction accuracy and computational costs.
- •For data platform operators: Use detailed decision frameworks to dynamically manage model updates or expansions in operational AI workflows based on evidence of failure types and costs.
Authors
Weihan Li, Xinlei Chen, Junhao Wu, Tianshi Zheng
Abstract
Learning systems adapt quickly inside a familiar family of models. The harder step comes earlier: deciding, from observations that could be noise, an exception, a change within the family or structure outside it, whether opening a richer family is worth its cost. We treat this as a costly sequential decision: prediction failure must be turned into structural evidence, evidence into a value of expansion, and value into action. The Structural Revision Environment produces matched failures from each source, varies the price of expansion and the remaining horizon independently of the evidence, and admits exact Bayesian calculations and an exact normative solution of the one-shot decision. Its solution shows that revision is a value boundary and not an evidence threshold: one history has different optimal actions under different prices, horizons and announced queries, the boundary between local repair and expansion is set by the inputs a rule predicts and a repair cannot cover, and belief in the richer family crosses long before the decision does. Transformers trained in the environment reproduce this boundary from utility alone. Language models of three post-training lineages carry a failure-sensitive signal in their predictions that is not reflected in their revision decisions, and given the gain of expanding they read it without weighing it against price and horizon. Three models allowed to reason weigh the stated gain in the reference's proportions and still do not turn the history into an estimate of what expansion would buy. Controlled post-training of the meta-trained learners moves the prior and the sharpness of predictions, and neither moves the criterion.