[arXiv]score: 0.15
LRP-Based Head Selection Improves Function Vector Efficiency and Accuracy in LLMs
June 4, 2026
Gradient-based attention head selection via Layer-wise Relevance Propagation substantially improves both efficiency and accuracy of function vectors for LLM steering, while distributed FV application outperforms simple aggregation. Code is publicly available.
cs.CLcs.LG
HOW THIS AFFECTS YOU
●
builderPublic code means you can directly test improved FV-based steering in existing LLM pipelines today.
●
researcherOffers concrete design choices for function vector formulation that improve task-steering performance without architectural changes.