🧪 Test?View on arXiv
LLMs Know the Constraint But Do Not Use It: Activation Bottlenecks in Pragmatic Constraint Reasoning
Author1, Author2, Author3, Author4, Author5
reasoningconstraint activationLLMsmodel evaluation
2608.12321
Builder Relevance
6h ago70%
Abstract
The paper explores how LLMs fail to utilize implicit feasibility constraints due to routing issues rather than a lack of knowledge.
Reality Card
Core Claim
The study identifies that hidden-constraint failure in LLMs is primarily a routing problem, not a knowledge problem, revealing that constraints are encoded but not effectively activated in decision-making.
Method / Result
Activation patching repairs one model's performance by +6.4 nats, while another shows no improvement (-0.07).
Limitations
The findings indicate that all prompted interventions inflate conservative bias, suggesting a limitation in generalizability across different models.