Search papers, labs, and topics across Lattice.
The paper introduces IH-Challenge, a reinforcement learning dataset designed to improve instruction hierarchy (IH) robustness in large language models by addressing challenges like confounding IH failures with instruction-following failures and learning shortcuts. Fine-tuning GPT-5-Mini on IH-Challenge, using online adversarial example generation, significantly improves IH robustness by 10% across various benchmarks and reduces unsafe behavior. The released dataset aims to facilitate further research on robust instruction hierarchy.
GPT-5-Mini can be made 10% more robust to jailbreaks and prompt injections simply by RL fine-tuning on a new instruction hierarchy dataset, IH-Challenge.
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.