Does weight-space programming need to touch the hidden layers?
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
An output-head experiment on Qwen3.5-0.8B edited only three output rows, leaving every transformer tensor unchanged, and scored 55.2% on MATH-500 against 52.2% for a published Astra-edited reference run that changed internal feed-forward layers.