Imprint Reader: From Weight-Update Readout to Behavioral Intervention
An Imprint Reader describes frozen weight updates and guides edits that raise refusal and tool-use scores.
The paper introduces the Imprint Reader, trained with Semantic Mount-and-Read Tuning (SaRT) to describe frozen language-model weight updates in natural language. On held-out updates, the joint Reader reaches judge-based Pass@100 of 2% for knowledge and 16% for behavior. At a 0.5% pruning rate, Reader-guided MetaEdit raises harmful-prompt refusal from 57.9% to 64.1%. Using behavior descriptions without target-task data, MetaEdit also lifts BFCL Overall from 41.69% to 44.60%.
- SaRT mounts each update and elicits a description with an anchor-free meta-query.
- Held-out Pass@100 is 2% for knowledge and 16% for behavior.
- 0.5% Reader-guided pruning lifts harmful-prompt refusal from 57.9% to 64.1%.
- MetaEdit raises BFCL Overall from 41.69% to 44.60% without target-task data.
From the original article
1 min read at huggingface.co
As language models take a growing role in AI development, a natural aspiration is for them to reflect on their own learning process, as humans do, and use that reflection to improve themselves. At the same time, these models have an advantage that human learners lack, since training leaves parameter-level traces that can, in principle, be inspected directly.
By Guanxu Chen, Qihao Lin, Jing Shao · 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.