LLMs struggle to decode mythological knowledge beyond dominant traditions

A new research paper investigates how 18 open-source large language models represent and decode mythological knowledge. The study found that while models can represent knowledge about deities like Zeus, Jupiter, and Thor, they struggle to consistently identify counterparts in less-represented mythologies such as Finnish, Slavic, Egyptian, or Chinese traditions. The research indicates that this failure occurs during the decoding or readout phase, not in the initial representation, and is influenced by the language used in the prompt. AI

IMPACT Highlights limitations in LLM cultural understanding and suggests avenues for improving cross-cultural knowledge representation.

RANK_REASON Academic paper detailing a novel methodology for analyzing LLM knowledge representation and decoding. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• activation patching
• arXiv
• Chinese mythology
• Egyptian mythology
• Finnish mythology
• Hugging Face
• Jupiter
• linear probing
• Logit Lens
• output extraction
• Slavic mythology
• Thompson-motif
• Thor
• Zeus

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →