LLMs Know Constraints but Fail to Use Them: Activation Bottlenecks in Pragmatic Reasoning

A recent study published on arXiv (submission 2608.12321) examines the reasons behind the failures of large language models (LLMs) when a prominent surface cue contradicts an implicit feasibility constraint. The researchers define this issue as 'conditional constraint activation,' differentiating between the model's internal understanding of the constraint and its application in decision-making. They introduce a four-part diagnostic that assesses knowledge (whether the constraint is encoded), symmetry (its equal encoding in both constraint-present and -absent prompts), routing (the use of the constraint in final decisions), and repair (the ability to rectify the decision through donor activation). Testing 14 models reveals two distinct failure modes. Probes on two open-weight models achieve over 88% accuracy in decoding the constraint, with activation patching improving one model (+6.4 nats) while harming the other (-0.07 nats). The study also investigates mitigation strategies, concluding that no prompted intervention successfully reaches the repair corner, as all interventions increase conservative bias through prerequisite mention. The authors assert that hidden-constraint failure is primarily a routing issue rather than a knowledge issue. This research falls under Computer Science > Computation and Language and includes references, citations, and related code and data.