In-Context Probing: Toward Building Robust Classifiers via Probing Large Language Models

TL;DR

This paper introduces In-Context Probing (ICP), a method that enhances the robustness of classifiers built on large language models by probing representations instead of decoding outputs, showing improved stability and efficiency especially with small datasets.

Contribution

The paper proposes In-Context Probing, a novel approach that improves classifier robustness and performance on small datasets by probing representations rather than decoding outputs.

Findings

ICP is more robust to instruction variations.

ICP performs comparably or better than fine-tuning.

ICP is effective with fewer than 100 training examples.

Abstract

Large language models are able to learn new tasks in context, where they are provided with instructions and a few annotated examples. However, the effectiveness of in-context learning is dependent on the provided context, and the performance on a downstream task can vary considerably, depending on the instruction. Importantly, such dependency on the context can surface in unpredictable ways, e.g., a seemingly more informative instruction might lead to a worse performance. In this paper, we propose an alternative approach, which we term In-Context Probing (ICP). Similar to in-context learning, we contextualize the representation of the input with an instruction, but instead of decoding the output prediction, we probe the contextualized representation to predict the label. Through a series of experiments on a diverse set of classification tasks, we show that in-context probing is significantly more robust to changes in instructions. We further show that ICP performs competitive or superior to finetuning and can be particularly helpful to build classifiers on top of smaller models, with less than a hundred training examples.