ReportGem ReportGem

Academic paper

Evaluation Pitfalls and Sparsity Limitations in LLM-based Confidence Estimates for Classification

Authors: Elena Merdjanovska, Omar Zaidan, Andreas R\"uckl\'ePublished: 2026-08-05Paper ID: 2608.04899Category: cs.CLLicense: CC BY 4.0

Abstract

Confidence estimation is essential when LLMs are used for classification, indicating when predictions can be trusted. However, common approaches such as verbalization produce extremely sparse outputs. For instance, Qwen3-32B verbalizes only eight unique confidence values on SST-2, with over half being exactly 95%, a pattern we observe consistently across four datasets and two LLMs. Besides limiting practical utility, we show that this sparsity critically affects evaluation: the choice of interpolation in area under the accuracy-rejection curve (AUARC) dramatically alters rankings, with consistency sampling dropping from best to worst under stepwise versus linear interpolation. We advocate for standardizing stepwise interpolation for a fairer comparison. Under such a fair evaluation, we find that weighting verbalized digits by token probabilities, a method we term verbalization logprobs, addresses sparsity and achieves the best AUARC (+2.3 points over vanilla verbalization) without incurring additional inference cost.

This public page contains bibliographic metadata and the author abstract. Use the reader for licensed document access.

Open licensed paper reader