activation-editing

Tag

Cards List
#activation-editing

Reference-Grafting Matches Fine-Tuning at Eliciting Sandbagged Capabilities

arXiv cs.LG · 2026-09-01 Cached

This paper introduces reference-grafting, a method to elicit sandbagged capabilities in AI models by editing activations, matching fine-tuning's effectiveness without weight updates or training labels.

0 favorites 0 likes
← Back to home

Submit Feedback