LLMs exhibit stronger illusory pattern perception than humans, study finds

A new research paper explores the phenomenon of illusory pattern perception in large language models (LLMs), finding that these models frequently exhibit stronger tendencies to infer meaningful relationships in random data than humans do. The study adapted psychological paradigms to LLM tasks, revealing that models over-associate positive attributes with majority groups and construct causal narratives from ambiguous events. Researchers developed a feature interpretability framework using Sparse Autoencoders to analyze internal representations, linking holistic frequency perception and analytic cognitive orientation to these illusory perceptions. The findings suggest a cognitive-like illusion that could impact the reliability of LLM reasoning. AI

IMPACT Highlights a potential cognitive-like illusion in LLMs that could affect their reasoning reliability and downstream applications.

RANK_REASON Research paper published on arXiv detailing a new finding about LLM behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• DagsHub
• Hugging Face
• illusory pattern perception
• large language models
• Sparse Autoencoders

• paper
• safety

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