Position: Uncertainty Quantification in LLMs is Just Unsupervised Clustering

In three linesA position paper critiques Uncertainty Quantification (UQ) methods for LLMs, arguing they are merely unsupervised clustering algorithms. These approaches quantify internal consistency of generations rather than external correctness, failing to detect 'confident hallucinations.' The author proposes a paradigm shift toward UQ grounded in objective truth.

Summary generated by Claude — human-verified