AI Paper Atlas

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

Evaluating LLM Uncertainty in Long-Form Generation Using Deterministic Ground Truth

Ido Amit, Ido Galil, Ran El-Yaniv

ICML 2026 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks · 0 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
add reasoning (CoT prompting or reasoning-specific training)paired Instruct/Reasoning models · SALTPrecision: 78.2%Improved
add reasoning (CoT prompting or reasoning-specific training)paired Instruct/Reasoning models · SALTECE: 60.8%Improved
add reasoning (CoT prompting or reasoning-specific training)paired Instruct/Reasoning models · SALTAUROC (ranking): -23.3%Hurt
add reasoning, controlling for precisionpaired Instruct/Reasoning models · SALTcalibration (ECE): p=0.567No clear effect
evaluate at atomic-unit instead of line-unit resolutionconfidence ranking · SALTMacro-AUROC: 0.45 to 0.7 (atomic Micro-AUROC)Hurt
use logits-based instead of verbalized confidence functionsconfidence functions · SALTcalibration and ranking: p < 0.05Improved
apply empirical binned calibration (vs Z-score+Sigmoid, Min-Max)post-hoc calibration · SALTECEImproved
apply empirical binned calibration (vs Z-score+Sigmoid, Min-Max)post-hoc calibration · SALTAUROC (ranking)Hurt
corrupt/fix prefix atoms (global prefix correctness, no local control)regenerated target atom · SALT DNA taskP(next correct) slope: +0.42 slope; +0.117 correction vs -0.231 corruptionImproved
corrupt/fix last-10 atoms (local prefix correctness, no global control)regenerated target atom · SALT DNA taskP(next correct) slope: +0.29 slope; +0.063 correction vs -0.238 corruptionImproved
corrupting prefix versus correcting itregenerated target atom · SALT DNA taskP(next correct): |corruption/correction| 2.0x (global), 3.8x (local)Hurt
remove prefix-correctness controls from length-response analysisanswer-context length curve · SALT DNA taskdrop in P(correct): -0.43 -> -0.50No clear effect

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