A single hypernetwork trained on 16,000 steering prompts can produce working activation steering vectors for new, unseen prompts, beating per-concept ReFT-r1 and rivaling prompting.
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
-
HyperSteer: Activation Steering at Scale with Hypernetworks
A single hypernetwork trained on 16,000 steering prompts can produce working activation steering vectors for new, unseen prompts, beating per-concept ReFT-r1 and rivaling prompting.