Anna Karenina Strikes Again: Pre-Trained LLM Embeddings May Favor High-Performing Learners

TL;DR

The paper interrogates whether unsupervised clustering of pre-trained LLM embeddings can recover educator-defined Knowledge Profiles (KPs) for open-ended biology responses. Using KMeans and HDBSCAN on Hebrew AlephBERT embeddings for 669 responses, it finds weak global alignment with KPs (low ARI) but relatively better retrieval for the KP representing correct answers, and reveals an Anna Karenina principle where high-quality responses form dense, central clusters while incorrect responses are diverse and less discoverable. The authors argue that out-of-the-box embeddings may be pedagogically insufficient for targeted formative feedback without domain-specific tuning or supervision, and they highlight biases toward correct responses as a risk in profile discovery. These findings call for careful evaluation of embedding-based profiling in education and motivate future work across languages and larger, more nuanced models to better support learners in need of personalized guidance.

Abstract

Unsupervised clustering of student responses to open-ended questions into behavioral and cognitive profiles using pre-trained LLM embeddings is an emerging technique, but little is known about how well this captures pedagogically meaningful information. We investigate this in the context of student responses to open-ended questions in biology, which were previously analyzed and clustered by experts into theory-driven Knowledge Profiles (KPs). Comparing these KPs to ones discovered by purely data-driven clustering techniques, we report poor discoverability of most KPs, except for the ones including the correct answers. We trace this "discoverability bias" to the representations of KPs in the pre-trained LLM embeddings space.