A new arXiv paper investigates how well large language models (LLMs) understand figurative language compared to humans. Researchers compared human participants and four LLMs—GPT-4, Gemma-2-9B, Llama-3.2, and Mistral-7B—on their interpretations of sentences involving conventionality, sarcasm, humor, emotion, idiomacy, and slang. While the models' surface-level judgments often aligned with humans, their underlying representational similarities differed, particularly for idiomatic expressions and slang. GPT-4 showed the closest alignment to human representations, but all models exhibited distinct internal processing patterns. AI
IMPACT This research highlights that while LLMs can mimic human understanding of figurative language, their internal representations may differ, suggesting potential limitations in nuanced comprehension.
RANK_REASON The cluster contains an academic paper detailing research findings on LLM capabilities. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Gemma
- Gemma-2-9B-IT
- Gemma-3-27B-IT
- Gemma-4-31B-IT
- GPT-4
- Hugging Face
- Llama~3.2
- mistral:7b
- Samhita Bollepally
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
