CLaRE study reports faster prediction of side effects from model edits

Researchers report that CLaRE, a technique for predicting unintended effects of changes to a language model's factual associations, improves average Spearman correlation with those effects by 62.2% against baselines. The preprint reports that the technique is 2.7x faster and uses 2.85x less peak GPU memory. CLaRE measures relationships between facts using activations from one intermediate model layer, avoiding backward passes used by gradient-based methods. The researchers analyzed 11,427 facts from three existing datasets and built graphs across multiple models. The method identifies where editing side effects may occur; it does not itself establish that an edit is safe.