Activation-based methods for selecting in-context examples don't work well in practice—the internal signals are too noisy to be useful, likely because models compress multiple features into the same dimensions.
This paper tests whether analyzing internal neural network activations can help select better examples for in-context learning in large language models. Despite thorough testing across multiple models and datasets, the researchers found that activation-based signals don't reliably predict example quality, suggesting this approach isn't practical for improving LLM few-shot performance.