Training vision models for LLM-driven weight-space programming
Original title: Training Vision Models for LLM-Driven Weight-Space Programming
Five LLMs ran more than 2,600 edits; Edit-Ready Training scored highest for GPT-6 Astra and Claude Opus 5.5.
The paper asks whether dedicated training makes vision networks easier for large language models to edit directly in weight space. Five LLMs attempt relabeling and error-correction edits on convolutional networks and vision transformers across Waterbirds, CIFAR-100, PACS and Entity-30. Baselines include standard training, label smoothing and gradient-based editable training. The authors plan to release the testing framework, tasks, trajectories and scores.
- Each edit is scored on held-out images for success, side effects and accuracy loss.
- ERT rehearses the direct, closed-form weight edits that LLMs perform.
- Preparing networks for editing raises success where standard networks are hard to edit.
- The project page is viscom.uni-ulm.de/wsp.
- AI models
- GPT-6 AstraClaude Opus 5.5
From the original article
1 min read at huggingface.coIt has been demonstrated that, in certain settings, large language models can perform weight-space programming to directly modify the weights of neural networks. However, it is unknown how dedicated training of these networks can contribute to the successes of weight-space programming.
By Timo Ropinski, Sebastian Kranz · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.