Tell, Don't Show: Internalized Reasoning influences how LLMs generalize

TL;DR — When training an LLM on examples that follow a pattern, but also giving the model declarative information stating the opposite of the pattern at train time, we investigate how the model ends up generalizing.

Abstract

In this paper we investigate to what extent language models' generalization behavior during a domain shift can be influenced by declarative knowledge contained in the training data. In order to study this we finetune language models to fit some distribution which has a ``natural'' generalization when the distribution shifts. We then test to what extent declarative statements in the training data - that if fully internalized would greatly affect the domain shift generalization - can indeed alter the model's behavior on unseen examples. While the effect is subtle, the declarative knowledge provided in the finetuning sets systematically changes the models' predictions in the way one would expect. Evidence for the strength of this effect growing with model size is mixed. We further show that the effect can not be explained by simple token matching behavior as it persists even when there is no overlap between the declarative descriptions and the models' test time generations.