Intel / article

Steering Instruction Hierarchies at Inference Time

Source: arxiv.org
Author: Siqi Zeng, Sewoong Lee, Han Zhao, Julia Hockenmaier
Date: Published Jul 30, 2026

Context

V-Steer tries to help a obey higher-priority instructions when user or tool text conflicts with them. Instead of retraining the model, it changes cached numerical information associated with the prompt: it identifies components where lower-priority text dominates, then strengthens the higher-priority material and weakens the conflicting material. The authors report primary-constraint accuracy rising from below 18% to as high as 92% in controlled role-conflict evaluations across models of different sizes. The intervention adds an initial prompt-processing cost, while subsequent text generation incurs negligible speed overhead in their tests. Method and reported results.

Key takeaways · AI-distilled

V-Steer restores system-prompt priority at time by editing the cached value vectors of the prompt: it multiplies up the privileged spans and damps the conflicting user or tool spans. No retraining, no weight changes.
It picks what to edit with direct logit attribution on the first predicted token. Attention heads where the low-priority span outweighs the privileged one are exactly the heads that get corrected.
On controlled role-conflict tests, obedience to the top-priority constraint rises from under 18% to 92%, holding across models from 7B to 70B parameters.
Because the edits touch only cached values, the method stays compatible with fused attention kernels and costs a one-time prefill hit instead of slowing down every token you decode.

Terms in this piece · Glossary

LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
system prompt — The standing instructions a model receives before any user input — defining its role, rules, tools, and tone for the whole conversation.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.

Key quotes

“Instruction hierarchies are a core safety assumption of language model deployment: higher priority inputs, such as system prompts, should override conflicting lower priority inputs from users or tools.”

“We introduce V-Steer, a training-free inference time method that restores privileged influence by editing cached value vectors at prompt positions.”

“Since the method acts only on cached values, it remains compatible with fused attention backends and adds only a one time prefill overhead.”

Read the source arxiv.org