NeurIPS Paper: Neural Networks Learn to Execute Algorithmic Instructions Exactly

kfountou · x · 2026-08-17

A NeurIPS 2025 paper investigates whether neural networks can learn to execute binary-encoded algorithmic instructions exactly using the Neural Tangent Kernel (NTK) framework. The study demonstrates that by structuring training data to isolate bit-level rules and controlling correlations in the NTK regime, two-layer fully connected networks can achieve high-probability exact execution of binary permutations, addition, multiplication, and SBN instructions. As SBN is Turing-complete, this framework extends to computable functions.

Original post →

More from Research

Research channel →