Steering Instruction Hierarchies at Inference Time

• URL: http://arxiv.org/abs/2607.26228v1
• Date: Tue, 28 Jul 2026 20:06:31 GMT
• Title: Steering Instruction Hierarchies at Inference Time
• Abstract summary: We introduce V-Steer, a training-free inference time method that restores privileged influence by editing cached value vectors at prompt positions.<n>V-Steer identifies heads where lower priority spans dominate privileged ones, then boosts privileged spans and suppresses conflicting lower priority spans.<n>Across models from 7B to 70B, this intervention raises primary constraint accuracy from under 18% up to 92% on controlled role conflict benchmarks.
• Score: 12.819290639427445
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: 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. Yet frontier LLMs often violate this hierarchy. We introduce V-Steer, a training-free inference time method that restores privileged influence by editing cached value vectors at prompt positions. Using direct logit attribution on the first next token prediction, V-Steer identifies heads where lower priority spans dominate privileged ones, then boosts privileged spans and suppresses conflicting lower priority spans through in-place multiplicative edits to cached V tensors. Since the method acts only on cached values, it remains compatible with fused attention backends and adds only a one time prefill overhead. Across models from 7B to 70B, this attribution guided intervention raises primary constraint accuracy from under 18% up to 92% on controlled role conflict benchmarks, and on broader instruction hierarchy evaluations substantially outperforms prompt only baselines while matching or exceeding SoTA training based methods on 3 of 4 scales of LLMs, with negligible decoding-speed overhead. The code is available at https://github.com/cindy2000sh/v-steer.

Related papers

• IH-Benchmark: A Conflict-Centered Benchmark for Instruction-Hierarchy Robustness in LLM Applications [1.9966419762834917]
  IH-Benchmark is a conflict-centered benchmark for instruction-hierarchy robustness across direct system-user conflicts (S>U) and tool-mediated user-tool (U>T) conflicts.<n>Across 37 evaluated models, hierarchy compliance ranges from 98.2% to 20.5%.<n>We find that strong S>U compliance is not a reliable proxy for U>T robustness.
  arXiv  Detail & Related papers  (2026-07-28T17:06:13Z)
• TypedCSIP: Typed Counterfactual Pretraining for Chinese Legislative Conflict Classification [3.5137191090796054]
  TypedCSIP is a typed counterfactual pretraining method for the conflict-classification task of the LCR-CN benchmark.<n>We exploit LCR-CN's expert-written minimal revisions as training-time counterfactual supervision.
  arXiv  Detail & Related papers  (2026-05-25T06:26:46Z)
• Verifier-Free RL for LLMs via Intrinsic Gradient-Norm Reward [69.99652051809737]
  In this work, we propose Verifier-free Intrinsic Gradient-Norm Reward (VIGOR)<n>VIGOR is a simple reward that uses only the policy model itself.<n>It exhibits cross-domain transfer to code benchmarks when trained only on math data.
  arXiv  Detail & Related papers  (2026-05-11T03:15:37Z)
• Reasoning Up the Instruction Ladder for Controllable Language Models [26.068755167791505]
  Large language model (LLM) based systems take on high-stakes roles in real-world decision-making.<n>Enforcement of an instruction hierarchy (IH) in LLMs is critical for the reliability and controllability of LLMs.<n>In this work, we reframe instruction hierarchy resolution as a reasoning task.
  arXiv  Detail & Related papers  (2025-10-30T22:13:31Z)
• Command-V: Pasting LLM Behaviors via Activation Profiles [67.07238260037839]
  Command-V is a backpropagation-free behavior transfer method.<n>It copies an existing residual activation adapter from a donor model and pastes its effect into a recipient model.
  arXiv  Detail & Related papers  (2025-06-23T21:21:49Z)
• PrePrompt: Predictive prompting for class incremental learning [37.06693481489317]
  Class Incremental Learning (CIL) based on pre-trained models offers a promising direction for open-world continual learning.<n>We propose Predictive Prompting (PrePrompt), a novel CIL framework that circumvents correlation-based limitations by leveraging pre-trained models' natural classification ability to predict task-specific prompts.<n> Experiments across multiple benchmarks demonstrate PrePrompt's superiority over state-of-the-art prompt-based CIL methods.
  arXiv  Detail & Related papers  (2025-05-13T13:57:56Z)
• Mitigating Word Bias in Zero-shot Prompt-based Classifiers [55.60306377044225]
  We show that matching class priors correlates strongly with the oracle upper bound performance. We also demonstrate large consistent performance gains for prompt settings over a range of NLP tasks.
  arXiv  Detail & Related papers  (2023-09-10T10:57:41Z)
• Task Residual for Tuning Vision-Language Models [69.22958802711017]
  We propose a new efficient tuning approach for vision-language models (VLMs) named Task Residual Tuning (TaskRes) TaskRes explicitly decouples the prior knowledge of the pre-trained models and new knowledge regarding a target task. The proposed TaskRes is simple yet effective, which significantly outperforms previous methods on 11 benchmark datasets.
  arXiv  Detail & Related papers  (2022-11-18T15:09:03Z)
• Prototypical Classifier for Robust Class-Imbalanced Learning [64.96088324684683]
  We propose textitPrototypical, which does not require fitting additional parameters given the embedding network. Prototypical produces balanced and comparable predictions for all classes even though the training set is class-imbalanced. We test our method on CIFAR-10LT, CIFAR-100LT and Webvision datasets, observing that Prototypical obtains substaintial improvements compared with state of the arts.
  arXiv  Detail & Related papers  (2021-10-22T01:55:01Z)
• Pre-training Is (Almost) All You Need: An Application to Commonsense Reasoning [61.32992639292889]
  Fine-tuning of pre-trained transformer models has become the standard approach for solving common NLP tasks. We introduce a new scoring method that casts a plausibility ranking task in a full-text format. We show that our method provides a much more stable training phase across random restarts.
  arXiv  Detail & Related papers  (2020-04-29T10:54:40Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.