Research map / Large language models / Language, safety and interpretability
Uncertainty quantification for LLMs
Quantify confidence and uncertainty to benchmark and evaluate language models.
136 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2019–2026) · an approach in Language, safety and interpretability, keyword cluster “safety · jailbreak · harmful”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Leveling off Several papers from 2025–26 still report on 1 tracked benchmark, but the best result has barely moved since 2024.
26 papers in 2023–24 and 109 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 3.7% to 5.9%.
What looks settled
- TriviaQA (AUROC): best up to 2024 0.866, best in 2025–26 0.845 from 3 papers, at most 0.2% better, if at all.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2019 · Cognitive model priors for predicting human decisions (ICML 2019 · 35 citations)
- 2023 · Prompting GPT-3 To Be Reliable (ICLR 2023 · 67 citations)
- 2024 · Membership Inference Attacks against Fine-tuned Large Language Models via Self-prompt Calibration (NeurIPS 2024 · 28 citations)
- 2025 · Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge (ICLR 2025 · 9 citations)
- 2026 · Not All Inconsistency Is Equal: Decomposing LVLM Uncertainty into Belief Divergence and Belief Conflict (AAAI 2026 · 2 citations)
Earliest papers here
- Cognitive model priors for predicting human decisions (ICML 2019 · 35 citations)
- Prompting GPT-3 To Be Reliable (ICLR 2023 · 67 citations)
- Membership Inference Attacks against Fine-tuned Large Language Models via Self-prompt Calibration (NeurIPS 2024 · 28 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
TriviaQA · AUROC (higher is better)
5 papers, 2024–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 0.866 | Semantic Density: Uncertainty Quantification for Large Language Models through Confidence Measurement in Semantic Space (NeurIPS 2024 · 6 citations) |
What has been tried
64 papers here report 706 measured changes (a component added, removed or swapped, compared with a baseline): helped 64%, no clear effect 8%, hurt 21%, mixed 8%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add practical difficulty calibration (PDC) | Wiki, GPT-J and LLaMA · AUC | 0.648 -> 0.929; 0.653 -> 0.951 | Helped | Membership Inference Attacks against Fine-tuned Large Language Models via Self-prompt Calibration (NeurIPS 2024 · 28 citations) |
| Add prompt covariates to PE-Rasch | MMLU, BBH, and LMentry · Wasserstein-1 distance | Helped | Efficient multi-prompt evaluation of LLMs (NeurIPS 2024 · 8 citations) | |
| Use logprob differences instead of binary yes/no features | Out-of-distribution lie detection · AUC | Helped | How to Catch an AI Liar: Lie Detection in Black-Box LLMs by Asking Unrelated Questions (ICLR 2024 · 7 citations) | |
| Reduce the number of reference responses from 10 toward 1 | CoQA, TriviaQA, SciQ, and NQ · AUROC | Hurt | Semantic Density: Uncertainty Quantification for Large Language Models through Confidence Measurement in Semantic Space (NeurIPS 2024 · 6 citations) | |
| Add confidence sampling rather than returning a single randomly sampled response | StrategyQA · accuracy | 49.20 -> 53.10 | Helped | Who Knows the Answer? Finding the Best Model and Prompt for Each Query Using Confidence-Based Search (AAAI 2024 · 5 citations) |
| use heuristic rather than random masking during corruptor training | FECDATA test, retrieved evidence · SARI Final | 60.04 -> 65.59; 58.54 -> 61.30 | Helped | Improving Factual Error Correction by Learning to Inject Factual Errors (AAAI 2024 · 4 citations) |
| Replace frequency with semantic diversity as the reliability measure | VMME · APSS | 2.81 -> 2.74; 2.58 -> 2.52; 2.70 -> 2.42; 2.24 -> 2.20 | Helped | Sample then Identify: A General Framework for Risk Control and Assessment in Multimodal Large Language Models (ICLR 2025 · 3 citations) |
| Add test-time feature clipping | CoQA · AUROC (sentence similarity) | 79.3 -> 80.4 | Helped | INSIDE: LLMs' Internal States Retain the Power of Hallucination Detection (ICLR 2024 · 3 citations) |
| Use KLE(K_heat) instead of KLE(K_full) | Llama 2 70B Chat, Trivia QA · AUROC | 0.80 ± 0.03 -> 0.84 ± 0.03 | Helped | Kernel Language Entropy: Fine-grained Uncertainty Quantification for LLMs from Semantic Similarities (NeurIPS 2024 · 3 citations) |
| Add maximum probability statistic | WikiBio sentence-level NonFact · AUC | 86.48 -> 90.93 | Helped | Enhancing Uncertainty Modeling with Semantic Graph for Hallucination Detection (AAAI 2025 · 2 citations) |
| Vary the number of sampled layers | uncertainty estimation across evaluated models · ECE | Mixed | FLUE: Streamlined Uncertainty Estimation for Large Language Models (AAAI 2025 · 2 citations) | |
| Use reversal-augmented backdoor data rather than backdoored data without augmentation | Backdoor trigger elicitation · Trigger elicitation rate | approximately 30% | Helped | Tell me about yourself: LLMs are aware of their learned behaviors (ICLR 2025 · 2 citations) |
What reviewers will expect
Datasets and baselines named in the 54 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- MMLU · 10 of 54 papers
- TriviaQA · 6 of 54 papers
- TruthfulQA · 6 of 54 papers
- GSM8K · 5 of 54 papers
- CommonsenseQA · 4 of 54 papers
- MATH · 4 of 54 papers
- OpenbookQA · 3 of 54 papers
- LiveBench · 2 of 54 papers
Baselines
- P(True) · 4 of 54 papers
- Semantic Entropy · 3 of 54 papers
- PandaLM-7B · 2 of 54 papers
- JudgeLM-7B · 2 of 54 papers
- GPT-4o · 2 of 54 papers
- LAC · 2 of 54 papers
- Perplexity · 2 of 54 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 135 papers; from OpenAlex and the papers' affiliations).
- Stanford University · 6 papers
- Microsoft · 6 papers
- National University of Singapore · 5 papers
- IBM · 4 papers
- University of Cambridge · 4 papers
- Amazon · 4 papers
- Massachusetts Institute of Technology · 3 papers
- University of North Carolina at Chapel Hill · 3 papers
- Apple · 3 papers
- University of Oxford · 3 papers
Use it
Download BibTeX of 22 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Statistical frameworks for LLM evaluation (124 papers)
- Bias measurement and mitigation (59 papers)
- Jailbreak elicitation and risk (21 papers)
Elsewhere in the topic
- LLM unlearning and benchmarks (298 papers)
- LLM pruning and scaling data (285 papers)
- Controlled text generation (220 papers)
- Faithfulness of LLM explanations (169 papers)
- Data and evaluation methodology (136 papers)
- Pruning and circuits in LLMs (135 papers)
- Safety alignment of LLMs (123 papers)
- Linguistic encoding in LLM representations (123 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).