Imprint Reader: From Weight-Update Readout to Behavioral Intervention
The paper tests whether a model can read frozen weight updates as language, then use that readout to steer behavior.
The authors introduce Imprint Reader, trained with Semantic Mount-and-Read Tuning to describe what a weight update appears to encode. On held-out updates, its judge-based Pass@100 is still low: 2% for knowledge and 16% for behavior. They also use the Reader as a proxy signal for editing, reporting higher harmful-prompt refusal at a 0.5% pruning rate and gains on BFCL Overall. Source: HF Daily Papers' note.
The authors introduce Imprint Reader, trained with Semantic Mount-and-Read Tuning to describe what a weight update appears to encode. On held-out updates, its judge-based Pass@100 is still low: 2% for knowledge and 16% for behavior. They also use the Reader as a proxy signal for editing, reporting higher harmful-prompt refusal at a 0.5% pruning rate and gains on BFCL Overall. Source: HF Daily Papers' note.
score 4