A CPU Inside an LLM: Zero-Gradient Compilation of Exact Circuits into Pretrained Weights
Researchers introduced closed-form, zero-gradient circuit-grafting primitives that compile deterministic computation directly into pretrained language model weights without requiring training examples or runtime hooks. The approach combines Mellin-GLU quadrature, cross-token scale cancellation, and axis-aligned width extension to synthesize arithmetic and logic circuits in constant depth. Evaluations show that models like IBM Granite and Alibaba Qwen achieve near-perfect ALU-prefix accuracy on multi-digit arithmetic while preserving natural-language competence. Open article →