In-Context Probing: Toward Building Robust Classifiers via Probing Large Language Models
This addresses the issue of unpredictable performance variations in instruction-based classifiers for users of large language models, offering a more reliable alternative.
The paper tackles the problem of in-context learning's sensitivity to instructions by proposing In-Context Probing (ICP), which probes contextualized representations instead of decoding predictions, resulting in significantly improved robustness and competitive or superior performance to finetuning with fewer than a hundred examples.
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.