The paper introduces NLS depth, a metric to quantify unverbalized serial cognition by measuring the longest computational path that avoids interpretable bottlenecks. It defines natural-language-rooted nodes to ensure the measure is grounded in human-understandable outputs.
HOW THIS AFFECTS YOU
●
researcherYou can use NLS depth to predict how much a model's reasoning will be hidden from chain-of-thought monitoring.