attribution-guided

Tag

Cards List
#attribution-guided

Steering Instruction Hierarchies at Inference Time

arXiv cs.CL · 2026-07-30 Cached

Introduces V-Steer, a training-free inference-time method that edits cached value vectors to restore instruction hierarchy in language models, raising primary constraint accuracy from under 18% to 92% on controlled benchmarks with negligible overhead.

0 favorites 0 likes
← Back to home

Submit Feedback