Cryptanalytic Extraction of Bias-Free GLU Feed-Forward Blocks Achieved

A novel cryptanalytic technique has been introduced for extracting isolated, bias-free Gated Linear Unit (GLU) feed-forward blocks from neural networks, addressing a gap in previous methodologies. This technique, outlined in a paper on arXiv (2608.06631), utilizes a multi-stage forward-query recovery primitive that is constructive in nature. It identifies candidates for gate direction through finite-difference curvature and distinguishes gate magnitude, orientation, and value-branch coupling by pairing observations at x and -x. The method was tested on high-precision targets, including six Qwen layers, an 8,192-unit Llama subproblem, and a complete Gemma block, all achieving a median validation error of less than one percent. This advancement builds on earlier extraction methods that were confined to ReLU networks and certain activation functions, specifically focusing on the two-branch architecture of GLU blocks. The paper details the algorithm and experimental findings, representing a major leap in model interpretability and security assessment.