LLMs Can Directly Read and Understand Neural Network Weights: Arcturus Labs Explores Frontier Models' Abstract Generalization

Frontier models have demonstrated the ability to directly infer and reverse-engineer neural network behaviors without training loops, using floating-point weights solely as plain text inputs. Tests showed that when given a semantically unlabelled network comprising 569 variables and 17,184 connection weights, models such as GPT-6 Astra and Opus 5.5 quickly identified that it corresponded to Pac-Man maze corridors and moving entity mechanics. This phenomenon indicates that model capabilities in program tracing and reinforcement learning generalization have expanded to the structured interpretation of complex numerical matrices.