ACL Findings2026
Evaluation Pitfalls and Sparsity Limitations in LLM-based Confidence Estimates for Classification
Elena Merdjanovska, Omar Zaidan, Andreas Rücklé
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.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
- Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs cites Large Language Models are Zero-Shot Reasoners
- Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs cites Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Generation
- Kernel Language Entropy: Fine-grained Uncertainty Quantification for LLMs from Semantic Similarities cites Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Generation
Builds on19
- Large Language Models are Zero-Shot ReasonersTakeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo et al.NeurIPS 2022 · 8,168 citations
- Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMsMiao Xiong, Zhiyuan Hu, Xinyang Lu, Yifei Li et al.ICLR 2024 · 867 citations
- Kernel Language Entropy: Fine-grained Uncertainty Quantification for LLMs from Semantic SimilaritiesAlexander Nikitin, Jannik Kossen, Yarin Gal, Pekka MarttinenNeurIPS 2024 · 197 citations
- Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language GenerationLorenz Kuhn, Yarin Gal, Sebastian FarquharICLR 2023 · 49 citations
- Prometheus 2: An Open Source Language Model Specialized in Evaluating Other Language ModelsSeungone Kim, Juyoung Suk, Shayne Longpre, Bill Yuchen Lin et al.EMNLP 2024 · 38 citations
Related papers
- How do LLMs Compute Verbal Confidence?Dharshan Kumaran, Arthur Conmy, Federico Barbero, Simon Osindero et al.ICML 2026 · 16 citations
- Confidence Improves Self-Consistency in LLMsAmir Taubenfeld, Tom Sheffer, Eran Ofek, Amir Feder et al.ACL Findings 2025 · 13 citations
- “very likely” Means “uncertain”? How LLMs Diverge from Humans in Linguistic Uncertainty QuantificationJinhao Duan, Zicheng Liu, Zijie Liu, Kaidi Xu et al.ICML 2026
- CoCoA: A Minimum Bayes Risk Framework Bridging Confidence and Consistency for Uncertainty Quantification in LLMsRoman Vashurin, Maiya Goloburda, Albina Ilina, Aleksandr Rubashevskii et al.NeurIPS 2025 · 33 citations
- Contextualized Sequence Likelihood: Enhanced Confidence Scores for Natural Language GenerationZhen Lin, Shubhendu Trivedi, Jimeng SunEMNLP 2024 · 1 citation
