Challenges with unsupervised LLM knowledge discovery

TL;DR

This paper critically examines unsupervised methods for discovering knowledge in large language models, revealing they often identify prominent features rather than actual knowledge, and provides theoretical and experimental evidence of this limitation.

Contribution

The paper demonstrates that existing unsupervised knowledge discovery methods are insufficient and introduces sanity checks for evaluating future approaches.

Findings

Unsupervised methods often detect prominent features, not knowledge.

Theoretically, arbitrary features satisfy the consistency structure used in these methods.

Experiments show classifiers predict features other than true knowledge.

Abstract

We show that existing unsupervised methods on large language model (LLM) activations do not discover knowledge -- instead they seem to discover whatever feature of the activations is most prominent. The idea behind unsupervised knowledge elicitation is that knowledge satisfies a consistency structure, which can be used to discover knowledge. We first prove theoretically that arbitrary features (not just knowledge) satisfy the consistency structure of a particular leading unsupervised knowledge-elicitation method, contrast-consistent search (Burns et al. - arXiv:2212.03827). We then present a series of experiments showing settings in which unsupervised methods result in classifiers that do not predict knowledge, but instead predict a different prominent feature. We conclude that existing unsupervised methods for discovering latent knowledge are insufficient, and we contribute sanity checks to apply to evaluating future knowledge elicitation methods. Conceptually, we hypothesise that the identification issues explored here, e.g. distinguishing a model's knowledge from that of a simulated character's, will persist for future unsupervised methods.