[Paper Review] Larger language models do in-context learning differently

Jerry Wei, Jason Lee+9 more|arXiv (Cornell University)|Mar 7, 2023

Topic Modeling99 citations

TL;DR

The paper shows that in-context learning relies on semantic priors in smaller models but emerges as an ability to learn input–label mappings in larger models, including under semantically unrelated labels and flipped exemplars, with instruction tuning further shaping these abilities.

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.

Motivation & Objective

• Investigate how semantic priors from pretraining influence in-context learning (ICL) across model sizes.
• Examine whether larger models can override semantic priors using in-context input–label mappings.
• Study ICL when labels are semantically unrelated to inputs (SUL-ICL) to test learning of input–label mappings.
• Assess the impact of instruction tuning on ICL, semantic priors, and input–label mapping learning.
• Evaluate emergent abilities in high-dimensional tasks such as linear classification under SUL-ICL.

Proposed method

• Compare regular ICL, flipped-label ICL, and semantically-unrelated label ICL (SUL-ICL) across multiple model families and sizes.
• Use in-context exemplars (k=16 per class by default) and evaluate over diverse NLP tasks with held-out labels.
• Systematically flip in-context labels to test override of semantic priors by large models.
• Replace natural language targets with semantically unrelated tokens (Foo/Bar) to force learning of input–label mappings.
• Assess impacts of instruction tuning (Flan-PaLM) versus pretrained models on ICL, priors, and mappings.
• Include high-dimensional linear classification tasks to probe non-language ICL capabilities.

Experimental results

Research questions

• RQ1 Can small language models override semantic priors with in-context input–label mappings when exemplars are flipped?
• RQ2 Do larger language models acquire the ability to learn input–label mappings in-context even when labels have no semantic relation to the task (SUL-ICL)?
• RQ3 How does instruction tuning affect reliance on semantic priors versus learning input–label mappings in ICL?
• RQ4 Is the ability to perform high-dimensional linear classification emergent with model scale under SUL-ICL?

Key findings

• Large models can override semantic priors when presented with flipped in-context labels, whereas small models largely cannot.
• Under SUL-ICL, model performance improves with scale, indicating emergent ability to learn input–label mappings without semantic priors.
• Instruction-tuned models better learn input–label mappings but also strengthen semantic priors, reducing ability to override priors with flipped labels.
• Performance gains from additional exemplars are larger for big models in SUL-ICL settings, suggesting better use of in-context mappings by large models.
• Some tasks show emergence of SUL-ICL capabilities only at large model scales (e.g., certain RTE and ETHOS tasks).
• Large models can even perform linear classification in high-dimensional settings under SUL-ICL.