Do LLM Embedding Spaces Recover Expert Structure?

This paper investigates whether the geometric structure of Large Language Model (LLM) embeddings—specifically the distances between category prototypes—actually reflects expert-defined conceptual knowledge or merely coarse distributional patterns. Using mental health-related Reddit communities as a testbed, the authors compare pretrained and fine-tuned Qwen3 embeddings (0.6B and 4B parameters) to determine if these models organize categories according to symptom overlap.

The researchers employ Representational Similarity Analysis (RSA) to compare model-derived dissimilarity matrices against an expert-curated symptom matrix. To ensure the results are not driven by non-clinical factors, they use prototype-based typicality and multi-baseline controls to account for affective valence, arousal, dominance (VAD), linguistic style (LIWC), and topic distributions.

The study demonstrates that LLM embeddings are not just black-box classifiers; they possess an underlying geometry that aligns with expert-defined symptom structures. Pretrained models show measurable alignment, but supervised fine-tuning acts as a catalyst, significantly sharpening the representation of fine-grained category boundaries.

Crucially, the authors find that model scale matters: larger models (4B) consistently outperform smaller ones (0.6B) in both zero-shot alignment and the degree of improvement gained through fine-tuning. The analysis also reveals that supervision does not simply blur all boundaries; rather, it suppresses off-structure ambiguity while concentrating residual uncertainty among categories that are genuinely related in the expert symptom framework.

As researchers increasingly use LLM embeddings as representational maps for downstream analysis, it is vital to understand whether these spaces are semantically grounded or merely reflecting superficial discourse cues. This work provides a rigorous framework for validating embedding geometry against external expert references. It suggests that while LLMs can capture complex, expert-relevant structures, developers and researchers should test for these properties explicitly rather than assuming that high classification accuracy implies a high-fidelity recovery of conceptual organization.