New framework reveals controllable personality traits within LLMs

Researchers have developed a new framework to explore personality-related representations within Large Language Models (LLMs), adapting Funder's personality triad of person, situation, and behavior. This framework uses SAE decomposition to identify internal features associated with personality traits, validating their relevance through behavioral effects and activation patterns. Interventions on these features demonstrate bidirectional, cross-situational shifts in trait-related behavior, mirroring findings from human personality research and suggesting LLMs possess controllable, trait-like internal representations. AI

IMPACT This research provides a novel method for understanding and potentially controlling internal representations within LLMs, offering insights into their behavioral patterns.

RANK_REASON The item is a research paper detailing a new framework for analyzing LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• CatalyzeX
• Connected Papers
• DagsHub
• Funder
• Gotit.pub
• Hugging Face
• Litmaps
• LLMs
• SAE decomposition
• ScienceCast
• scite Smart Citations

• paper
• model release

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