New research identifies key factors for actionable latent structures in language models

Researchers have identified three key factors that determine whether a latent structure within a language model's activation space is actionable for controlling its behavior. These factors are capacity, which measures output sensitivity to movement along the structure; responsiveness, indicating how promotable a concept is within a given context; and alignment, reflecting how well the structure matches the context-specific representation of a concept. The study found that all three factors must be high for causal effectiveness, with low capacity and responsiveness significantly reducing it, and low alignment potentially reversing it. This research also suggests that causality is context-dependent, leading to the development of 'causal probes' that improve model steering capabilities. AI

IMPACT This research offers a framework for better understanding and controlling language model behavior, potentially leading to more reliable and steerable AI systems.

RANK_REASON The cluster contains a research paper published on arXiv detailing findings about language model behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• Connected Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• Influence Flower
• Language Models
• Litmaps
• ScienceCast
• scite Smart Citations

• paper
• model release

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