In merged language models, task-vector interference is causally carried by the orientation of an internal cross-term, not by its magnitude, and instruction wrappers can hide this interference while the model still carries it.
Zhanpeng Zhou, Zijun Chen, Yilan Chen, Bo Zhang, and Junchi Yan
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Orientation, not magnitude: the causal structure of task-vector interference in merged language models
In merged language models, task-vector interference is causally carried by the orientation of an internal cross-term, not by its magnitude, and instruction wrappers can hide this interference while the model still carries it.