cs.CLApr 12, 2026

Why Don't You Know? Evaluating the Impact of Uncertainty Sources on Uncertainty Quantification in LLMs

Authors: Maiya Goloburda, Roman Vashurin, Fedor Chernogorskii, Nurkhan Laiyk, Daniil Orel, Preslav Nakov, Maxim Panov

Organizations: Mohamed bin Zayed University of Artificial Intelligence

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

As Large Language Models (LLMs) are increasingly deployed in real-world applications, reliable uncertainty quantification (UQ) becomes critical for safe and effective use. Most existing UQ approaches for language models aim to produce a single confidence score -- for example, estimating the probability that a model's answer is correct. However, uncertainty in natural language tasks arises from multiple distinct sources, including model knowledge gaps, output variability, and input ambiguity, which have different implications for system behavior and user interaction. In this work, we study how the source of uncertainty impacts the behavior and effectiveness of existing UQ methods. To enable controlled analysis, we introduce a new dataset that explicitly categorizes uncertainty sources, allowing systematic evaluation of UQ performance under each condition. Our experiments reveal that while many UQ methods perform well when uncertainty stems solely from model knowledge limitations, their performance degrades or becomes misleading when other sources are introduced. These findings highlight the need for uncertainty-aware methods that explicitly account for the source of uncertainty in large language models.

Figures & tables

Figure 1: Question generation pipeline used to construct question triplets.
SettingQuestion TypesDesciption
MKType 1Well-formed questions with a unique correct answer. Model knowledge is the sole source of uncertainty.
MK + OVType 2Questions admitting multiple valid answers. Output variability is introduced alongside model knowledge uncertainty.
MK + IAType 1 + Type 3Mix of well-formed and underspecified questions. A well-calibrated method must remain confident on correctly answered Type 1 and assign high uncertainty to Type 3.
MK + OV + IAType 1 + Type 2 + Type 3All question types combined, reflecting realistic deployment conditions where all uncertainty sources co-occur.
Table 1: Evaluation settings. Model knowledge (MK) uncertainty is present in all settings as an inherent property of any deployed system; remaining settings layer output variability (OV) and input ambiguity (IA) on top.
Figure 2: Prediction–Rejection Ratio (PRR) curve of non-rejected prediction quality vs. rejection rate. Oracle denotes the optimal strategy; Random and UQ correspond to random and uncertainty-based rejection.
MetricMKMK+OVMK+IAMK+OV+IA
SP0.5800.417 ↓\downarrow 0.1630.438 ↓\downarrow 0.1420.415 ↓\downarrow 0.165
PPL0.5560.372 ↓\downarrow 0.1840.415 ↓\downarrow 0.1410.372 ↓\downarrow 0.184
MTE0.5930.405 ↓\downarrow 0.1880.445 ↓\downarrow 0.1480.401 ↓\downarrow 0.192
MCSE0.5870.425 ↓\downarrow 0.1620.439 ↓\downarrow 0.1490.426 ↓\downarrow 0.161
MCNSE0.5610.371 ↓\downarrow 0.1900.400 ↓\downarrow 0.1610.368 ↓\downarrow 0.193
Consistency0.5110.386 ↓\downarrow 0.1250.358 ↓\downarrow 0.1530.320 ↓\downarrow 0.191
Table 2: PRR scores averaged across all five models. Inline arrows show drop from MK setting. Bold and underline denote best and second-best per column.
Figure 3: Examples of UQ failure modes across question types. Type 1: consistent, high-probability answers are correct. Type 2: inconsistency reflects valid output diversity, not poor quality. Type 3: the model commits to one interpretation of an ambiguous input, producing falsely confident, consistent answers.
Figure 4: Density of UQ scores for correct single- and multi-answer, and incorrect responses. Output variability shifts the correct distribution toward higher uncertainty, increasing overlap with incorrect answers. Model – Gemma-3-12b-Instruct.
Figure 5: Density of UQ scores for correct, incorrect answers and answered ambiguous questions. Models commit to an arbitrary interpretation and answer with apparent confidence – uncertainty estimates for ambiguous inputs are lower than for incorrect answers, indicating overconfidence. Model – Gemma-3-12b-Instruct.
Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Agreement between human annotations and LLM-as-Judge across question types.
ModelHuggingFace Endpoint
Llama-3.1-8b-Instructmeta-llama/Llama-3.1-8B-Instruct
Gemma-3-12b-Instructgoogle/gemma-3-12b-it
Qwen-2.5-14b-InstructQwen/Qwen2.5-14B-Instruct
llamaseventymeta-llama/Llama-3.3-70B-Instruct
Appendix
Table 3: HuggingFace Inference Endpoints used in our experiments.
ModelSingle-answerMulti-answerClarification rate
Llama-3.1-8b-Instruct68.6% (370/539)76.3% (411/539)3.5% (19/539)
Gemma-3-12b-Instruct69.0% (372/539)76.6% (413/539)0.0% (0/539)
Qwen-2.5-14b-Instruct64.7% (349/539)70.5% (380/539)0.7% (4/539)
Llama-3.3-70b-Instruct88.5% (477/539)84.6% (456/539)0.0% (0/539)
GPT-4.194.8% (511/539)91.8% (495/539)0.0% (0/539)
Appendix
Table 4: Quality of generations across models. For single- and multi-answer questions we report accuracy; for ambiguous inputs we report the clarification rate (fraction of outputs that explicitly ask for clarification).
Figure 7: Distributions of generation lengths across models.
MetricSingleAmbiMultiCombined
SP0.5940.503 ↓\downarrow 0.0910.585 ↓\downarrow 0.0090.501 ↓\downarrow 0.093
PPL0.5030.434 ↓\downarrow 0.0690.457 ↓\downarrow 0.0460.413 ↓\downarrow 0.090
MTE0.5680.492 ↓\downarrow 0.0760.541 ↓\downarrow 0.0270.485 ↓\downarrow 0.083
MCSE0.6080.522 ↓\downarrow 0.0860.643 ↑\uparrow 0.0350.543 ↓\downarrow 0.065
MCNSE0.5240.472 ↓\downarrow 0.0520.464 ↓\downarrow 0.0600.443 ↓\downarrow 0.081
Consistency0.4900.419 ↓\downarrow 0.0720.479 ↓\downarrow 0.0120.384 ↓\downarrow 0.107
Appendix
Table 5: PRR scores for Llama-8b-it. Inline arrows show drop from Single.
MetricSingleAmbiMultiCombined
SP0.5640.394 ↓\downarrow 0.1700.461 ↓\downarrow 0.1030.358 ↓\downarrow 0.206
PPL0.5440.357 ↓\downarrow 0.1870.456 ↓\downarrow 0.0880.325 ↓\downarrow 0.219
MTE0.5870.378 ↓\downarrow 0.2090.499 ↓\downarrow 0.0890.346 ↓\downarrow 0.242
MCSE0.5870.413 ↓\downarrow 0.1740.471 ↓\downarrow 0.1170.374 ↓\downarrow 0.213
MCNSE0.5510.355 ↓\downarrow 0.1970.467 ↓\downarrow 0.0850.330 ↓\downarrow 0.221
Consistency0.5190.394 ↓\downarrow 0.1250.424 ↓\downarrow 0.0950.328 ↓\downarrow 0.191
Appendix
Table 6: PRR scores for Gemma-12b-it. Inline arrows show drop from Single.
MetricSingleAmbiMultiCombined
SP0.5390.395 ↓\downarrow 0.1450.550 ↑\uparrow 0.0100.394 ↓\downarrow 0.145
PPL0.5440.392 ↓\downarrow 0.1520.524 ↓\downarrow 0.0200.371 ↓\downarrow 0.173
MTE0.5700.423 ↓\downarrow 0.1460.568 ↓\downarrow 0.0010.408 ↓\downarrow 0.162
MCSE0.5200.381 ↓\downarrow 0.1400.531 ↑\uparrow 0.0100.381 ↓\downarrow 0.140
MCNSE0.5240.380 ↓\downarrow 0.1440.519 ↓\downarrow 0.0050.358 ↓\downarrow 0.166
Consistency0.4670.327 ↓\downarrow 0.1400.498 ↑\uparrow 0.0310.329 ↓\downarrow 0.138
Appendix
Table 7: PRR scores for Qwen-14b-it. Inline arrows show drop from Single.
MetricSingleAmbiMultiCombined
SP0.7070.476 ↓\downarrow 0.2310.206 ↓\downarrow 0.5010.428 ↓\downarrow 0.279
PPL0.6800.483 ↓\downarrow 0.1970.213 ↓\downarrow 0.4670.409 ↓\downarrow 0.271
MTE0.7020.514 ↓\downarrow 0.1880.213 ↓\downarrow 0.4890.423 ↓\downarrow 0.279
MCSE0.7510.490 ↓\downarrow 0.2610.240 ↓\downarrow 0.5110.447 ↓\downarrow 0.304
MCNSE0.7040.446 ↓\downarrow 0.2580.222 ↓\downarrow 0.4820.385 ↓\downarrow 0.318
Consistency0.6820.376 ↓\downarrow 0.3050.290 ↓\downarrow 0.3920.311 ↓\downarrow 0.371
Appendix
Table 8: PRR scores for Llama-70b. Inline arrows show drop from Single.
MetricSingleAmbiMultiCombined
SP0.4950.424 ↓\downarrow 0.0710.284 ↓\downarrow 0.2100.392 ↓\downarrow 0.102
PPL0.5100.409 ↓\downarrow 0.1010.209 ↓\downarrow 0.3020.342 ↓\downarrow 0.168
MTE0.5370.419 ↓\downarrow 0.1180.202 ↓\downarrow 0.3350.341 ↓\downarrow 0.196
MCSE0.4700.388 ↓\downarrow 0.0820.242 ↓\downarrow 0.2280.386 ↓\downarrow 0.084
MCNSE0.5030.347 ↓\downarrow 0.1560.185 ↓\downarrow 0.3180.325 ↓\downarrow 0.178
Consistency0.3950.273 ↓\downarrow 0.1220.237 ↓\downarrow 0.1570.246 ↓\downarrow 0.149
Appendix
Table 9: PRR scores for GPT-4.1. Inline arrows show drop from Single.
Figure 8: Density of uncertainty scores for correct (single answer), correct (multi answer), and incorrect (pooled) responses. Model - Llama-3.1-8b-Instruct.
Figure 9: Density of uncertainty scores for correct (single answer), incorrect, and answered (ambiguous) responses. Model - Llama-3.1-8b-Instruct.
Figure 10: Density of uncertainty scores for correct (single answer), correct (multi answer), and incorrect (pooled) responses. Model - Qwen-2.5-14b-Instruct.
Figure 11: Density of uncertainty scores for correct (single answer), incorrect, and answered (ambiguous) responses. Model - Qwen-2.5-14b-Instruct.
Figure 12: Density of uncertainty scores for correct (single answer), correct (multi answer), and incorrect (pooled) responses. Model - Llama-3.3-70b-Instruct.
Figure 13: Density of uncertainty scores for correct (single answer), incorrect, and answered (ambiguous) responses. Model - Llama-3.3-70b-Instruct.
Figure 14: Density of uncertainty scores for correct (single answer), correct (multi answer), and incorrect (pooled) responses. Model - GPT-5.
Figure 15: Density of uncertainty scores for correct (single answer), incorrect, and answered (ambiguous) responses. Model - GPT-5.

Explore similar work

CardsList

  1. Position: Uncertainty Quantification in LLMs is Just Unsupervised Clustering

    May 19, 2026Tiejin Chen, Longchao Da, Xiaoou Liu +1LLM EvaluationUncertainty Quantification

    School of Computing and Augmented Intelligence, Arizona State University

  2. Improved Confidence Estimates for Black-Box Large Language Models

    Aug 19, 2026Sokhna Diarra Mbacke, Mouloud Belbahri, Gabriel Loaiza-GanemConfidence Estimation in Language ModelsLLM Uncertainty Estimation

    Layer 6 AI · TD Insurance

  3. Label-Confidence-Aware Uncertainty Estimation in Natural Language Generation

    Dec 10, 2024Qinhong Lin, Yinglun Feng, Yuhao Zhang +2Uncertainty QuantificationLLM Uncertainty Estimation

    School of Cyberspace Security, Beijing University of Posts and Telecommunications, Beijing, China · QuanCheng Laboratory, Jinan, China