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

2026-08-05

AI Summary

This work addresses the unreliability of confidence estimation in large language models (LLMs) for classification tasks when using verbalized outputs, which often suffer from sparsity. The authors systematically analyze how this sparsity adversely affects evaluation metrics such as AUARC and propose a standardized evaluation protocol based on stepwise interpolation to ensure fair comparisons. They further introduce a token log-probability–weighted verbalization confidence method—termed *verbalization logprobs*—that mitigates output sparsity without incurring additional inference overhead. Experimental results demonstrate that this approach improves AUARC by 2.3 percentage points over baseline methods under the standardized stepwise interpolation protocol, substantially enhancing the quality of confidence estimates.

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.

Problem

• confidence estimation
• sparsity
• LLM-based classification
• AUARC evaluation
• verbalization

Innovation

• confidence estimation
• verbalization sparsity
• AUARC evaluation
• stepwise interpolation
• verbalization logprobs