Beyond Fixed Architectures: An LLM Agent Rewires Trained Neural Networks with Exact, Tested Patches
Code, patches, git histories, models and paper. Gradient descent changes the values of a network's weights inside an architecture fixed before training. Here an LLM coding agent (Claude Opus 5.5, directed by a person) rewires small trained networks after training instead. It writes patches out of exact identities of the (Leaky)ReLU activations the networks already use, which carry, gate, bound and switch signals, and makes room for them by freeing channels by regression or by growing channels with zero outgoing weights. On four small models the patches do what training the model does not: a coordinate wire and an AND gate give a cat GAN odd-e
Resource information
- Format
- zip
- Original repository
- 23985135
- DOI
- 10.5281/zenodo.23985135
- Credit basis
- Creators in Zenodo record metadata
- Access
- Open files available from the original Zenodo deposit
- Last checked
- Nov 5, 2026
- License
- mit-license
- Last record update
- Nov 5, 2026
About this resource
Code, patches, git histories, models and paper. Gradient descent changes the values of a network's weights inside an architecture fixed before training. Here an LLM coding agent (Claude Opus 5.5, directed by a person) rewires small trained networks after training instead. It writes patches out of exact identities of the (Leaky)ReLU activations the networks already use, which carry, gate, bound and switch signals, and makes room for them by freeing channels by regression or by growing channels with zero outgoing weights. On four small models the patches do what training the model does not: a coordinate wire and an AND gate give a cat GAN odd-eyed cats while the other half of every image moves by at most 1.6e-7; a reference path and a min-gate repair a 2018 super-sampler with an output bound that holds on every input; a self-checking router resolves a fine-tune merge conflict (+0.23 / +3.26 / +0.00 dB, against +0.03 / +3.19 / +0.08 for joint retraining); and one gated unit fixes 22 of 23 MobileNetV2 cat errors while changing 1 of 1,000 other predictions. Patches are hash-pinned, self-testing text files that wpgit builds, tests in context, merge-checks and rebases under git; the repository's name, WSSE, stands for weight-space software engineering. Paper: https://v-code-895.github.io/wsse/