rather than simply putting it into the token embedding, the limitation is that it manipulates the residual stream (activation) at a specific position to inject it
Latent Interpretation Tuning
fine-tune a LLM with activation–question–answer triples
LatentQA
Creator
Creator
Seonglae ChoCreated
Created
2025 Dec 19 12:13Editor
Editor
Seonglae ChoEdited
Edited
2025 Dec 19 12:21Refs
Refs