LLMs Display Alien Cognitive Structures, Study Finds
Recent findings published on arXiv question the belief that AI and humans possess similar cognitive frameworks. Researchers examined answers from humans and six large language models (LLMs) on assessments in quantitative reasoning and chemistry through Exploratory Factor Analysis. Experts in the field evaluated the resulting factor graphs without prior knowledge of their origins. They were able to interpret the majority of factors from human responses, but struggled to assign meaning to any factors derived from LLMs in quantitative reasoning, and only half in chemistry. These results imply that LLMs might depend on unfamiliar latent structures that do not correspond to human-understandable concepts.
Key facts
- Study conducted on arXiv preprint 2608.17810
- Investigates latent factors in LLM performance
- Uses responses from humans and six LLMs
- Assessments cover quantitative reasoning and chemistry
- Methodology: Exploratory Factor Analysis (EFA) separately for humans and LLMs
- Subject-Matter Experts blindly evaluated factor graphs
- SMEs interpreted most human-derived factors
- SMEs could not interpret any LLM factors in quantitative reasoning, and only half in chemistry
Entities
—