LLMs Know the Constraint But Do Not Use It: Activation Bottlenecks in Pragmatic Constraint Reasoning
The episode discusses a paper by Carnegie Mellon researchers Yubo Li, Ramayya Krishnan, and Rema Padman. They find that Large Language Models (LLMs) possess knowledge of constraints but fail to apply them during decision-making. The hosts conclude that this is a routing problem, not a knowledge deficit, and requires architectural fixes like activation patching rather than prompting.