Larger language models do in-context learning differently

TL;DR

This paper investigates how larger language models perform in-context learning by examining their ability to override semantic priors and learn input-label mappings, revealing that these capabilities emerge primarily with increased scale and are enhanced by instruction tuning.

Contribution

The study demonstrates that the ability to override semantic priors and learn input-label mappings in in-context learning emerges mainly with larger models and is improved by instruction tuning.

Findings

Large models can override semantic priors with contradictory in-context exemplars.

Large models can perform linear classification in semantically-unrelated label settings.

Instruction tuning enhances both semantic prior use and input-label mapping learning.

Abstract

We study how in-context learning (ICL) in language models is affected by semantic priors versus input-label mappings. We investigate two setups-ICL with flipped labels and ICL with semantically-unrelated labels-across various model families (GPT-3, InstructGPT, Codex, PaLM, and Flan-PaLM). First, experiments on ICL with flipped labels show that overriding semantic priors is an emergent ability of model scale. While small language models ignore flipped labels presented in-context and thus rely primarily on semantic priors from pretraining, large models can override semantic priors when presented with in-context exemplars that contradict priors, despite the stronger semantic priors that larger models may hold. We next study semantically-unrelated label ICL (SUL-ICL), in which labels are semantically unrelated to their inputs (e.g., foo/bar instead of negative/positive), thereby forcing language models to learn the input-label mappings shown in in-context exemplars in order to perform the task. The ability to do SUL-ICL also emerges primarily with scale, and large-enough language models can even perform linear classification in a SUL-ICL setting. Finally, we evaluate instruction-tuned models and find that instruction tuning strengthens both the use of semantic priors and the capacity to learn input-label mappings, but more of the former.