Ulm Benchmark Shows Edit-Ready Training Lets LLMs Reliably Rewrite Vision Model Weights
Summary
Ulm researchers released a benchmark where five frontier LLMs edit convolutional networks and vision transformers on Waterbirds, CIFAR-100, PACS and Entity-30 relabeling and error-correction tasks across 2,600+ edit trajectories. Their Edit-Ready Training (ERT) method rehearses the closed-form weight edits LLMs perform, substantially raising edit success where standard networks are hard to edit. ERT achieves the top mean score for GPT-6 Astra and Claude Opus 5.5 across all baselines.
Originally reported by huggingface.co
Read the original article →Original headline: Ulm Benchmark Shows Edit-Ready Training Lets LLMs Reliably Rewrite Vision Model Weights