New research reveals how language models choose between recall and copying

Researchers have developed a method to analyze how language models decide between using their internal knowledge or information from a provided context. They identified an "arbitration vector" that can be manipulated to either encourage the model to recall its parametric knowledge or to copy information directly from the prompt. Experiments across different model architectures and question-answering benchmarks demonstrated that inducing copying is a simpler process than suppressing it and restoring recall, which appears more sensitive to specific interventions. AI

IMPACT Provides insights into controlling language model behavior in retrieval-augmented settings, potentially improving accuracy and fluency.

RANK_REASON Research paper published on arXiv detailing a mechanistic study of language model behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• Mehrdad Farahani

• paper
• model release

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