Multi-Tier Instruction Hierarchy for LLM Agents

LLM agents simultaneously receive commands from multiple sources: system messages, user prompts, tool outputs, and instructions from other agents. When these conflict, the agent must accurately determine which instruction has higher authority in order to behave safely and usefully.

Existing instruction hierarchy (IH) approaches typically rely on fewer than 5 fixed authority levels and rigid role labels like system > user. But in real agent environments, conflicts arise from far more sources and contexts, so such simple hierarchies fall short.

To address this, the authors propose the Many-Tier Instruction Hierarchy (ManyIH). The core idea is to handle authority levels as arbitrarily many tiers rather than a limited few, enabling agents to reliably follow higher-authority instructions even in complex conflict situations.

Alongside this, they release ManyIH-Bench, the first benchmark for evaluating this problem. The benchmark includes conflicting instructions across up to 12 tiers and consists of a total of 853 agent tasks.

• 427 coding tasks
• 426 instruction-following tasks
• 46 tests reflecting real agent environments
• Constraints generated by LLMs were refined through human verification to increase realism and difficulty

Experimental results showed that even current frontier models suffered significant performance drops as instruction conflicts grew more complex, with accuracy reaching only about 40%. The authors emphasize that for agents to handle complex instruction conflicts, more fine-grained and scalable methods for interpreting authority are needed.