Thunder-NUBench: A Benchmark for LLMs' Sentence-Level Negation Understanding

  • 2025-06-18 02:18:11
  • Yeonkyoung So, Gyuseong Lee, Sungmok Jung, Joonhak Lee, JiA Kang, Sangho Kim, Jaejin Lee
  • 0

Abstract

Negation is a fundamental linguistic phenomenon that poses persistentchallenges for Large Language Models (LLMs), particularly in tasks requiringdeep semantic understanding. Existing benchmarks often treat negation as a sidecase within broader tasks like natural language inference, resulting in a lackof benchmarks that exclusively target negation understanding. In this work, weintroduce Thunder-NUBench, a novel benchmark explicitly designed to assesssentence-level negation understanding in LLMs. Thunder-NUBench goes beyondsurface-level cue detection by contrasting standard negation with structurallydiverse alternatives such as local negation, contradiction, and paraphrase. Thebenchmark consists of manually curated sentence-negation pairs and amultiple-choice dataset that enables in-depth evaluation of models' negationunderstanding.

 

Quick Read (beta)

loading the full paper ...