Steering Instruction Hierarchies at Inference Time

Short summary

V-Steer is a training-free, inference-time method that enforces instruction hierarchies in LLMs by editing cached value vectors at prompt positions. It identifies attention heads where lower-priority inputs dominate system prompts via direct logit attribution, then boosts privileged spans and suppresses conflicting ones with multiplicative edits to cached V tensors. Across 7B–70B models it raises primary constraint accuracy from under 18% to 92% on role-conflict benchmarks, matching or exceeding training-based methods with negligible decoding overhead.

• •V-Steer edits cached value vectors at inference time to enforce instruction hierarchy (system prompt over user/tool inputs)
• •Training-free, compatible with fused attention backends, adds only one-time prefill overhead
• •Raises constraint accuracy from <18% to 92% on role-conflict benchmarks across 7B–70B models; code open-sourced

Generated with AI, which can make mistakes.