A new arXiv paper tackles a core safety problem for deployed language models: higher-priority instructions, such as system prompts, should override conflicting user or tool text. The authors say frontier models still often violate this hierarchy.
Their method, V-Steer, works at inference time without training. It uses direct logit attribution on the first next-token prediction to find attention heads where lower-priority spans dominate privileged instructions. It then edits cached value tensors to boost the higher-priority spans and suppress conflicting lower-priority ones.
Because the method changes cached values, the authors say it remains compatible with fused attention backends and adds only a one-time prefill overhead. Across models from 7B to 70B parameters, it raised primary constraint accuracy from under 18 percent to as high as 92 percent on controlled role-conflict benchmarks.
The result is a research claim, not a finished platform feature. It points to a direction where inference-time controls may complement training and prompt design for safer instruction following.