Where Is Uncertainty Observed? A Mechanistic Understanding of Uncertainty Quantification in LLMs
Organizations: University College London · Technische Universität Darmstadt · Universität zu Lübeck · Mohamed bin Zayed University of Artificial Intelligence
Abstract
Knowing when a large language model's (LLM's) response can be trusted is critical for its safe use. Uncertainty quantification (UQ) can help LLMs recognize if their responses may be unreliable. Existing taxonomies organize UQ methods by uncertainty source or estimation mechanism. They leave unresolved the question of where uncertainty is observed in the model's computation. The answer matters because the observation point determines whether an estimator has access to enough information to recover a given target, and how well it can recover it. Drawing on mechanistic interpretability, we define four mechanistic observation lenses: the structural lens observes component updates, the representational lens observes hidden states, the distributional lens observes token probabilities, and the semantical lens observes distributions over meanings. We relate these lenses through the residual stream and order them by the information they carry, which gives a principle for selecting UQ methods for different targets. We survey existing methods through the four lenses. Furthermore, we distinguish evaluation targets, criteria, and metrics to clarify different UQ evaluation protocols. Finally, we identify open problems and research directions to connect uncertainty estimation to the internal computation and downstream use of LLMs.
Figures & tables
| Symbol | Meaning |
| Input space, output space, hypothesis space, probability distribution | |
| The LLM, parameterized by | |
| , | Input and output sequence, with and as the token at position |
| Residual-stream state after block , of hidden dimension | |
| Update written to the residual stream by component at layer | |
| Residual-stream index, is the embedding output |
| Type | Conditions | Implementations | ||
| Cluster | Symmetric, transitive | Equivalence classes | Bidirectional NLI | |
| Graph | or | Sym. or asym. | Nodes in weighted graph | Graded NLI |
| Manifold | Symmetric | Outputs in metric space | Embedding similarity | |
| Report | Internal | — | Implicit | Verbalized confidence |
| Type | Method | Access | N/S | Target | Eval | ||
| Cluster | Semantic Entropy ( Kuhn et al., 2023 ) | COR | RNK | ||||
| SE Blackbox ( Farquhar et al., 2024 ) | COR | RNK | |||||
| NumSets ( Lin et al., 2024b ) | COR | RNK | |||||
| Alphabet Size ( McCabe et al., 2026 ) | COR | RNK DEC | |||||
| Self-Assessment ( Cao et al., 2026 ) | COR | RNK SCR | |||||
| SRE-UQ ( Vipulanandan et al., 2026 ) | COR | RNK |
| Table continued from previous page | |||||||
| Type | Method | Access | N/S | Target | Eval | ||
| Report | Verbalized Confidence ( Lin et al., 2022a ) | COR | CAL | ||||
| Knowledge Transfer ( Yang et al., 2024b ) | COR | CAL | |||||
| SaySelf ( Xu et al., 2024 ) | COR | CAL RNK | |||||
| Ling. Calib. ( Band et al., 2024 ) | COR | CAL | |||||
| CONQORD ( Tao et al., 2024 ) | COR | CAL | |||||
| Level | Method | Access | N/S | Target | Eval | ||
| Token | MSP ( Hendrycks & Gimpel, 2017 ) | COR | RNK | ||||
| LogTokU ( Ma et al., 2025a ) | COR | RNK | |||||
| U-LINE ( Vashurin et al., 2025c ) | COR | RNK | |||||
| Semantic Energy ( Ma et al., 2025b ) | COR | RNK | |||||
| LoFree CP ( Su et al., 2024 ) | COR | COV | |||||
| CAPE ( Jiang et al., 2023b ) | COR | CAL RNK SCR |
| Layer | Method | Access | N/S | Target | Eval | ||
| Fixed | Rel. Mahalanobis Dist. ( Ren et al., 2023a ) | COR | RNK | ||||
| U-Space ( Braun et al., 2026 ) | COR | RNK | |||||
| INSIDE ( Chen et al., 2024 ) | COR | RNK | |||||
| SemGrad ( Li et al., 2026a ) | COR | RNK | |||||
| P(IK) ( Kadavath et al., 2022 ) | COR | CAL | |||||
| Probe Reasoning ( Zhang et al., 2025a ) | COR | RNK CAL |
| Comp. | Method | Access | N/S | Target | Eval | ||
| MHA | LLM Check ( Sriramanan et al., 2024 ) | COR GRD | RNK DEC | ||||
| TOHA ( Bazarova et al., 2026 ) | GRD | RNK | |||||
| RFS Guard ( Liu et al., 2026b ) | COR | RNK DEC | |||||
| Attention Head Entropy ( Ostmeier et al., 2026 ) | COR | RNK CAL | |||||
| SAT Probe ( Yuksekgonul et al., 2024 ) | COR | RNK | |||||
| UQAC ( Li et al., 2025b ) | COR | CAL RNK |
| Target | Criterion | Metric Family | Ground Truth | # Papers |
| Correctness COR | dis, sem | RNK CAL SCR DEC REG COV | ✓ | 146 |
| Grounding GRD | dis, sem | RNK DEC | ✓ | 11 |
| Underspecification UNS | sem | RNK DEC | ✓ | 3 |
| Faithfulness FTH | sem | FTH | – | 6 |
| Format | Benchmarks |
| MCQA | MMLU(-Pro) ( Hendrycks et al., 2021a ; Wang et al., 2024c ) , MedQA ( Jin et al., 2021 ) , MedMCQA ( Pal et al., 2022 ) , GPQA ( Rein et al., 2024 ) , CommonsenseQA ( Talmor et al., 2019 ) , BoolQ ( Clark et al., 2019 ) , ARC ( Clark et al., 2018 ) , HellaSwag ( Zellers et al., 2019 ) , FEVER ( Thorne et al., 2018 ) , Winogrande ( Sakaguchi et al., 2021 ) , PIQA ( Bisk et al., 2020 ) , GLUE ( Wang et al., 2018 ) , BigBench ( Srivastava et al., 2023 ) , GRD SciQ ( Welbl et al., 2017 ) , CosmosQA ( Huang et al., 2019 ) , RACE ( Lai et al., 2017 ) , OpenBookQA ( Mihaylov et al., 2018 ) , HaluEval ( Li et al., 2023a ) |
| SFQA | TriviaQA ( Joshi et al., 2017 ) , NQ(-open) ( Kwiatkowski et al., 2019 ) , PopQA ( Mallen et al., 2023 ) , BioASQ ( Tsatsaronis et al., 2015 ) , TruthfulQA ( Lin et al., 2022b ) , SimpleQA ( Wei et al., 2024 ) , MS MARCO ( Bajaj et al., 2018 ) , WebQuestions ( Berant et al., 2013 ) , LAMA ( Petroni et al., 2019 ) , PubMedQA ( Jin et al., 2019 ) , MGSM ( Shi et al., 2023 ) , GRD CoQA ( Reddy et al., 2019 ) , HotpotQA ( Yang et al., 2018 ) , SQuAD ( Rajpurkar et al., 2016 ) , HaluEval ( Li et al., 2023a ) , UNS QuAC ( Choi et al., 2018 ) , SelfAware ( Yin et al., 2023 ) |
| LFQA | ELI5(-Category) ( Fan et al., 2019 ; Atakishiyev et al., 2024 ) , MedLFQA ( Jeong et al., 2024 ) , GRD Qasper ( Dasigi et al., 2021 ) , NarrativeQA ( Kočiský et al., 2018 ) |
| OEG | FActScore ( Min et al., 2023 ) , WikiBio GPT-3 ( Manakul et al., 2023 ) , Chatbot Arena ( Chiang et al., 2024 ) , FAVA ( Mishra et al., 2024 ) , HumanEval ( Chen et al., 2021 ) , MBPP ( Austin et al., 2021 ) , GRD MT-Bench ( Chen et al., 2026b ) |
| CG | GRD XSum ( Narayan et al., 2018 ) , SamSum ( Gliwa et al., 2019 ) , TL;DR ( Stiennon et al., 2020 ) , SummEval ( Fabbri et al., 2021 ) , FRANK ( Pagnoni et al., 2021 ) , CNN/DailyMail ( Hermann et al., 2015 ) , WMT ( Bojar et al., 2017 ; Barrault et al., 2019 ; Barrault et al., 2020 ) , HaluEval ( Li et al., 2023a ) , RAGTruth ( Niu et al., 2024 ) |
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Model | fraction | fraction | rank / dim |
| Qwen2.5-7B-Instruct | / | ||
| Qwen2.5-7B (base) | / | ||
| Llama-3.1-8B-Instruct | / |
Uncertainty Quantification
Uncertainty quantification estimates how reliable predictions and scientific inferences are, and how uncertainty propagates into decisions.
See all 699 papers