AI Agents Boost Algorithmic Reasoning to 84% by Splicing Code Into Model Architectures

A new AI training process allows coding agents to modify a neural network's structure in real-time, moving the resulting models beyond the design of standard transformers. By splicing code layers directly into the network during training, researchers saw algorithmic reasoning accuracy for nanochat rise from 55% to 84%, while training loss fell below the levels seen in ordinary training. The method also enables open-weight models to perform self-modifications, such as DeepSeek-V4.1-Flash, which analyzed its own activations to identify hidden states signaling incorrect answers and wrote a new code layer to predict these failures, increasing its AIME and HMMT 2026 scores from 64% to 71%.