AI Paper Atlas

Research map / Optimization / Implicit bias and generalization

From Flat Facts to Sharp Hallucinations: Detecting Stubborn Errors via Gradient Sensitivity

Liew Yee Zhing, Andrew Huey Ping Tan, Anwar P.P. Abdul Majeed

ICML 2026 · Optimization › Implicit bias and generalization · approach: Edge-of-stability dynamics · 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
use EPGS instead of Eigenscorehallucination detector · TriviaQA, Llama-3-8BAUROC: 0.7226 -> 0.8127Improved
use EPGS instead of Semantic Entropy on stubborn subsethallucination detector · SQuAD stubborn subset, Llama-2-7BAUROC: 0.5842 -> 0.7373Improved
use EPGS with SQuAD-F1 labels instead of best baselineEPGS · SQuAD, Llama-2-7B (SQuAD-F1 labels)AUROC: 0.7578 -> 0.7109Hurt
use EPGS instead of Eigenscore on TruthfulQAhallucination detector · TruthfulQAAUROC: 0.34 - 0.39 -> 0.65 - 0.69Improved
extract gradients from last transformer block instead of final layer norm (ablation)EPGS · SQuAD, Llama-3-8BAUROC: 0.5928 -> 0.7742Improved
extract gradients from last transformer block instead of LLM head (ablation)EPGS · SQuAD, Llama-3-8BAUROC: 0.5474 -> 0.7742Improved
extract gradients from last block instead of middle block (ablation)EPGS · SVAMP, Llama-2-7BAUROC: 0.8031 -> 0.8947Improved
extract gradients from last block instead of middle block (ablation)EPGS · SQuAD, Llama-2-7BAUROC: 0.6965 -> 0.6924Hurt
vary Gaussian noise magnitude from 0.001 to 0.1 (ablation)EPGS · SQuAD, Llama-2AUROC: 0.7847 -> 0.6924Mixed
use embedding noise instead of paraphrasing (ablation)EPGS · TriviaQA, Llama-2-7BAUROC: 0.6985 -> 0.7629Improved
use embedding noise instead of dropout (ablation)EPGS · TriviaQA, Llama-3-8BAUROC: 0.8028 -> 0.8127Improved
add NER-based entity masking (ablation)EPGS · NQ, Llama-2-7BAUROC: as much as 8.2% drop without maskImproved

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