AI Paper Atlas

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

Gradients with Respect to Semantics Preserving Embeddings Tell the Uncertainty of Large Language Model

Mingda Li, Rundong Lv, Xinyu Li, Weinan Zhang, Ting Liu

ICML 2026 · Large language models › Language, safety and interpretability · approach: Statistical frameworks for LLM evaluation

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 entropy re-weighting of token log-likelihoods (ablation)SemGrad Qwen3-4B · TriviaQAAUROC: 76.83 -> 80.40Improved
add entropy re-weighting of token log-likelihoods (ablation)SemGrad Llama3.1-8B · TriviaQAAUROC: 81.28 -> 84.72Improved
use semantic preserving token t* instead of last input token (ablation)SemGrad Qwen3-4B · TruthfulQAAUROC: 63.35 -> 69.06Improved
use semantic preserving token t* instead of last input token (ablation)SemGrad Qwen3-4B · SciQAUROC: 70.49 -> 72.20Improved
use l1 norm instead of l2 norm (ablation)SemGrad · SciQ/TriviaQA/TruthfulQAAUROC: 72.07 -> 72.20No clear effect
use top-half layer span instead of all layers 1:(L-1) (ablation)SemGrad Qwen3-4B · TruthfulQAAUROC: 66.60 -> 69.06Improved
use top-half layer span instead of single layer L-1 (ablation)SemGrad Llama3.1-8B · SciQAUROC: 74.64 -> 75.76Improved
use SemGrad instead of SAR on multi-answer dataSAR · TruthfulQAAUROC: +3.27 pointsImproved
use SemGrad instead of ExGrad on multi-answer dataExGrad · TruthfulQAAUROC: +6.82 pointsImproved
use SemGrad instead of ParaGrad on multi-answer dataParaGrad · TruthfulQAAUROC: +3.3Improved
use SemGrad instead of ParaGrad on single-answer dataParaGrad · TriviaQA Qwen3-4BAUROC: 82.02 -> 80.40Hurt
combine SemGrad and ParaGrad into HybridGradSemGrad · Average over 9 pairsAUROC: 75.75 -> 76.59Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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