AI Paper Atlas

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

Addressing Pitfalls in the Evaluation of Uncertainty Estimation Methods for Natural Language Generation

Mykyta Ielanskyi, Kajetan Schweighofer, Lukas Aichberger, Sepp Hochreiter

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Uncertainty quantification for LLMs · 1 citations (OpenAlex)

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
select correctness function adversarially for UE methods (ablation)UE methods (e.g. G-NLL, Min Token Log-Prob) · QA datasetsTop-3 frequency: G-NLL 0.375 -> 0.688; Min Token Log-Probability 0.125 -> 0.500Improved
marginalize over multiple LLM judges (SP-MoJI, 4 judges) instead of a single judge (ablation)UE evaluation · QA datasetsstd of performance estimator: SD up to 0.04 with one judge; twofold lower with 4 judgesImproved
add more judge calls beyond about 10 (ablation)SP-MoJI · QA datasetsstd of performance estimatorNo clear effect
apply length normalization to UE scoresPredictive Entropy / Semantic Entropy / Perplexity · QA, code, constrained textAUROC / EloHurt
apply length normalization to UE scoresPerplexity, Semantic Entropy · Perturbation detectionEloImproved
adjust token likelihoods with semantic importance (TokenSAR/SAR)SAR family · Multiple tasksEloHurt
evaluate pretrained instead of instruction-tuned modelsUE methods · Elo partitionsEloHurt

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