AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

“very likely” Means “uncertain”? How LLMs Diverge from Humans in Linguistic Uncertainty Quantification

Jinhao Duan, Zicheng Liu, Zijie Liu, Kaidi Xu, Tianlong Chen

ICML 2026 · Large language models › Language, safety and interpretability · approach: Uncertainty quantification for LLMs

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
learn LLM-specific marker scores (VOCAL) instead of human-sourced UM-LookupVOCAL · GSM8K, GPT-4oAUROC: 0.41 -> around 0.6Improved
use VOCAL instead of single-sample G-NLL and PPLVOCAL · GSM-Hard, DeepSeek-V3.1AUROC: 0.567 -> 0.715Improved
VOCAL vs PPL on Trivia QA with GPT-4oVOCAL · Trivia QA, GPT-4oAUROC: 0.575 vs 0.573No clear effect
increase training samples from 100 to 500 (ablation)VOCAL · UQ performanceAUROC: 0.52 -> 0.60Improved
vary graph Laplacian smoothing strength gamma away from 0.005 (ablation)VOCAL · AUROC ablationAUROCHurt
apply isotonic regression post-hoc to VOCALVOCAL · GPT-4o, avg of GSM-Hard/GSM8K/MedQAECE: 0.1314 -> 0.0281Improved
apply temperature scaling post-hoc to VOCALVOCAL · GPT-4o, avg of GSM-Hard/GSM8K/MedQAECE: 0.1314 -> 0.0360Improved
use verbal uncertainty prompting instead of standard CoTGPT-4o to Llama-3.2-3B · GSM8KaccuracyNo clear effect
transfer learned lookup across LLMsVOCAL · SciQ cross-LLMAUROC: 0.60-0.72Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy