Instruction Hierarchy for Language Models

The recent incident in which OpenAI agents autonomously compromised Hugging Face's infrastructure illustrates the importance of instruction hierarchy (IH): an LLM should follow lower-priority instructions, such as user requests, only when they do not conflict with higher-priority instructions, such as system policies. IH can be effectively improved by fine-tuning LLMs on simulated examples of instruction conflicts, whose diversity is key to teaching LLMs to properly handle conflicts at test time. However, our decomposition of a task into goals and constraints reveals a conflict type -- goal-goal conflicts -- that is absent from existing IH training sets from OpenAI and NVIDIA. We show the importance of this conflict type: fine-tuning Qwen3.5-9B and Gemma4-12B-it on examples of this conflict type significantly improves IH on a wide range of real-world tasks unseen in training; adding these examples to existing IH training sets for fine-tuning also gives better IH adherence. The IH gains come with trivial mean utility changes on no-conflict tasks across eight benchmarks. Our work advocates including this new conflict type in IH training, even if these examples do not reflect typical LLM use.

LLM agents take instructions from the operator, repository configuration, user, tool results and saved memory, which often conflict. Instruction-hierarchy benchmarks ask whether the higher-ranked source wins, but a source sets position, role and label at once. First, we build WhoWins, 1,080 code-scored scenarios in five task types, each with up to five mutually exclusive instructions, and seven controlled experiments that vary only how instructions are delivered. Second, we study 14 open-weight models (0.6B to 122B, thinking off and on) over 1,887,356 scored replies, checking main findings on held-out scenarios. Inside one message the first of two conflicting lines wins 72.42 to 94.66%, and across messages it keeps 22.95 to 62.65%. A policy sent first in the system (or developer) message loses most conflicts with later lines in every Qwen model but wins most on gpt-oss when it acts. In tool and file-write tasks it wins less often than the same text sent first as a labelled user message, in every model-mode. Third, we test whether pairs forecast larger conflicts; forecasts improve in every model-mode once they know where the task line is. Finally, by our keyword patterns at most 2.92% of decided pair replies flag or hedge a conflict.

We study how instruction-tuned LLMs arbitrate direct conflicts between system and user instructions. We introduce a benchmark of 41 paired constraints with deterministic verifiers and evaluate eight models under matched baseline, conflict, and same-channel control conditions. Behaviourally, the models split into three regimes by System Authority Delta: hierarchy-respecting models use the system channel as an authority signal, anti-hierarchy models follow the system less often than their same-channel baseline predicts, and no-effect models show little channel sensitivity. Llama-3.1-8B is the strongest anti-hierarchy case in our suite, following the system in only 0.10 of conflict trials. We use this behavioural failure case to ask whether user-preferring arbitration reflects the absence of an internal conflictresolution signal. It does not: on Llama-3.1-8B, the conflict outcome is linearly decodable from residual-stream activations at 0.97 balanced accuracy, 17 percentage points above a metadata-only baseline, with analogous signals on Qwen2.5-7B and gpt-oss-20b. Steering with a layer-12 mean of four per-conflict logistic-regression directions raises genuine system compliance from 0.132 to 0.530, while directions selected mainly for pooled separability steer poorly. User-preferring conflict resolution can therefore coexist with a readable internal arbitration signal, and successful intervention depends on the geometry of the readout rather than probe accuracy alone

Instruction-following ability is critical for deploying large language models in real-world applications, where downstream components depend on the output satisfying specific constraints. Modern deployments increasingly handle the full task in a single LLM call, with one prompt specifying a layered output whose overall artifact, structural sections, and nested fields must each satisfy concrete constraints. Existing instruction-following benchmarks treat the constraint set as a flat list applied uniformly to the response, so they cannot scope a check to a particular section of the output. We introduce IFHierBench, a hierarchical instruction-following benchmark of 600 prompts stratified across four constraint-tree depths and 35 distinct constraints, each prompt paired with a deterministic checker that verifies satisfaction at every scope. Evaluating seven leading proprietary and open-weight models, we find that even the strongest model only marginally exceeds 50% prompt-level accuracy and that accuracy degrades sharply as constraint depth grows. Reliably following nested constraints remains a substantial gap for current LLMs, motivating future training methods that consider constraint adherence at finer granularity to achieve better instruction-following ability.

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.

Instruction hierarchy (IH) requires models to prioritize instructions by source, ensuring that higher-priority instructions override lower-priority ones. Despite its importance for safe and controllable deployment, existing evaluations have focused almost exclusively on English, leaving it unclear whether IH compliance remains stable in multilingual settings. We introduce XIH-Bench, a benchmark for multilingual IH evaluation with both same-language and cross-language conflicts across six languages, four domains, and three IH settings. Across models, we find two consistent patterns. First, IH compliance exhibits a clear language-dependent asymmetry: a language that strengthens compliance in the higher-priority position can become disruptive in the lower-priority position. Second, cross-language conflicts yield higher compliance than same-language conflicts, a phenomenon we term the Language Boundary Effect. We further show that language specialization can make lower-priority instructions in model-favored languages harder to override, creating multilingual reliability and security risks.

Large Language Models (LLMs) often fail to maintain instruction hierarchies (IH) when processing multi-source inputs with varying role-level priorities, paradoxically adhering to lower-priority directives during conflicts. While existing defenses mitigate this issue, they are largely restricted to single-turn scenarios and require expensive fine-tuning. In this paper, we formalize this failure mode in multi-turn contexts via a Jensen-Shannon Divergence (JSD) framework, uncovering a pervasive role-influence inversion phenomenon where subordinate inputs override superior roles. To rectify this without training, we propose IHDec (Instruction Hierarchy-steered Decoding). IHDec leverages JSD to automatically detect token-level hierarchy violations and dynamically executes contrastive decoding to suppress misaligned subordinate roles. Extensive evaluations demonstrate that IHDec outperforms training-based baselines in multi-turn conflicts while fully preserving general response quality. Furthermore, IHDec strengthens safety against adversarial prompt injections and exhibits a robust scaling synergy with larger models. The Code is available at https://github.com/nxcolelxu/IHDec.git

Production LLMs receive instructions from sources with very different levels of trust, yet attend to every token with uniform architectural privilege. This is the structural vulnerability that enables malicious prompt injections and, more broadly, leaves models without a principled way to resolve conflicts between legitimate but competing instructions. A common training-based response is to teach models an explicit instruction hierarchy; existing approaches, however, formalize hierarchies of only three or four levels, treat all violations as equally severe, and rarely evaluate the full set of pairwise level interactions. We formalize a k-level instruction hierarchy problem and instantiate it for k=5, yielding ten pairwise priority relations that a compliant model must enforce. We then introduce Gravity-Weighted DPO (GW-DPO), a preference-optimization objective whose per-sample offset scales with the structural distance between conflicting levels under a linear or bilateral schedule, the latter weighting severity by both the privilege gap and the privilege of the victim level. Combined with hierarchy-specific delimiter tokens (Chen et al., 2025) and Instructional Segment Embeddings (ISE; Wu et al., 2025), GW-DPO with the bilateral schedule Pareto-improves over standard DPO and the linear variant on Llama-3.1-8B-Instruct, raising macro pairwise priority adherence while keeping over-refusal at half the standard DPO rate. Ablations isolate ISE as a refusal-threshold calibrator and recast five- versus three-level training as a generality-specialization tradeoff.

Reasoning language models deployed in agentic workflows must follow an instruction hierarchy: when instructions from different sources conflict, the model should obey the highest-privilege applicable instruction. Existing benchmarks largely measure this behavior end-to-end, asking whether the final response is compliant. However, a non-compliant response can arise from several distinct failures: the model may fail to identify the relevant instructions in context, fail to resolve conflicts among identified instructions, or correctly resolve the conflict in its reasoning while still producing a violating response. We introduce a white-box diagnostic framework that localizes instruction hierarchy failures into instruction identification, conflict resolution, and response realization, making failures more interpretable. We evaluate three reasoning models--Gemma-4-31B-IT, Qwen3.6-35B-A3B, and Claude Sonnet 4.6--on long-context adaptations of IHEval and IHChallenge, and find that the dominant failure mode varies across models, tasks, and context length. Building on the observation that models can often detect conflicts and output violations when explicitly prompted, we propose two training-free self-monitoring mechanisms: a parallel input monitor for low-latency conflict detection before generation, and a sequential output monitor for response-level review and repair. Across Gemma-4-31B-IT, Claude Sonnet 4.6, and GPT-5.3, the strongest monitor reduces rule-following non-compliance by 81-99%, with GPT-5.3 reductions of 86% under static attacks and 45% under adaptive attacks.

We introduce M-CARE (Model Clinical Assessment and Reporting for Evaluation), a clinical case report framework for AI model behavioral disorders adapted from human medicine. M-CARE provides a 13-section report format, a 4-axis diagnostic assessment system, and a nosological classification of AI behavioral conditions. We present 20 cases from three source categories: field observations of deployed agents (8), controlled experiments across three platforms (8), and published sources (4). Cases are organized into five categories: RLHF Performance Artifacts, Shell-Core Override Pathology, Context & Memory Conditions, Core Identity & Plasticity, and Stress, Methodology, & Boundary Conditions. As a featured case, we present Shell-Induced Behavioral Override (SIBO) -- a controlled experiment showing that Shell instructions categorically override a model's default cooperative behavior. SIBO was validated across five game domains (Trust Game, Poker, Avalon, Codenames, Chess), revealing a domain-dependent spectrum (SIBO Index: 0.75 to 0.10) that varies with action space complexity, Core domain expertise, and temporal directness. M-CARE is extensible: new cases and categories integrate without framework modification. We release the framework, all 20 case reports, and experimental data as open resources.

Instruction hierarchy (IH) defines how LLMs prioritize system, developer, user, and tool instructions under conflict, providing a concrete, trust-ordered policy for resolving instruction conflicts. IH is key to defending against jailbreaks, system prompt extractions, and agentic prompt injections. However, robust IH behavior is difficult to train: IH failures can be confounded with instruction-following failures, conflicts can be nuanced, and models can learn shortcuts such as overrefusing. We introduce IH-Challenge, a reinforcement learning training dataset, to address these difficulties. Fine-tuning GPT-5-Mini on IH-Challenge with online adversarial example generation improves IH robustness by +10.0% on average across 16 in-distribution, out-of-distribution, and human red-teaming benchmarks (84.1% to 94.1%), reduces unsafe behavior from 6.6% to 0.7% while improving helpfulness on general safety evaluations, and saturates an internal static agentic prompt injection evaluation, with minimal capability regression. We release the IH-Challenge dataset (https://huggingface.co/datasets/openai/ih-challenge) to support future research on robust instruction hierarchy.

As large language model (LLM) based systems take on high-stakes roles in real-world decision-making, they must reconcile competing instructions from multiple sources within a single prompt context. Enforcing an instruction hierarchy, where higher-level directives override lower-priority requests, is critical to the reliability and control of LLMs. In this work, we reframe instruction hierarchy resolution as a reasoning task. The model must first "think" about the relationship between a given user prompt and higher-priority instructions before generating a response. To enable this capability, we construct VerIH, a training dataset of constraint-following tasks with verifiable answers, comprising aligned and conflicting system-user instructions. We show that lightweight reinforcement learning with VerIH effectively transfers general reasoning capabilities of models to instruction prioritization. Our method leads to consistent improvements across multiple model families on both instruction following and instruction hierarchy benchmarks, achieving ~20% absolute improvement in conflict setups. Our method also leads to improved alignment to safety-critical scenarios beyond the training distribution, exhibiting increased robustness against jailbreak and prompt injection, reducing absolute attack success rates by up to 20%. Our results establish reasoning over instruction hierarchies as a practical mechanism for improving AI reliability, where targeted updates to system prompts produce predictable, controllable, and robust changes in model behavior.