AI & softwareResearch findings
Maximally Activating Example Metamodel for Residual Stream Direction Interpretation
Original published 2026.11.06 03:54 KSTAdded to Blackdeck 2026.11.08 23:33 KST
What has changed?
A metamodel that converts residual stream directions into text exemplars that strongly activate them is proposed.
Key points
- We propose a metamodel, MAEM, to interpret residual stream directions by generating text exemplars that strongly activate them.
- MAEM maximizes cosine similarity between a given residual direction and activations of generated text, enabling precise interpretation.
- MAEM outperforms corpus search and NLA in activating held-out SAE features and recovering LoRA backdoor payloads.
Connect it to your work
It maximizes cosine similarity between a given direction and generated text activations to produce informative examples.
What remains unproven
This research note is based on an abstract; check the original for test conditions and application scope.
Originals used for this note
Links open the original publisher in a new tab.