A large-scale study of task vectors in Llama-3-8B shows performance peaks at layer 15 and that complex tasks require multiple subtask-specific vectors rather than a single vector.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
One Task Vector is not Enough: A Large-Scale Study for In-Context Learning
A large-scale study of task vectors in Llama-3-8B shows performance peaks at layer 15 and that complex tasks require multiple subtask-specific vectors rather than a single vector.