Explain the limited utility of high-dimensional query embeddings for confidence estimation

Determine whether the lack of consistent improvements from augmenting uncertainty-score and neighbourhood-statistics features with high-dimensional query embeddings is caused by the dimensionality of the embeddings, which may require substantially more complex models and larger training sets to extract signals predictive of LLM-response correctness.

Background

The proposed confidence-estimation framework represents each LLM response using pre-existing uncertainty scores and statistics derived from similar queries in a labelled reference set. The authors also experimented with adding high-dimensional query embeddings to this feature representation, but did not observe consistent performance gains over models using uncertainty scores and neighbourhood statistics alone.

The paper offers a conjectural explanation for this result: the dimensionality of query embeddings may make it difficult for the comparatively simple classifiers and available training data to reliably learn features predictive of response correctness. Establishing whether this explanation is correct would clarify when query embeddings are useful for black-box LLM uncertainty quantification and what model capacity or dataset size is required to exploit them.