Reproduce “Many-Tier Instruction Hierarchy in LLM Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jingyu Zhang, Tianjian Li, William Jurayj, Hongyuan Zhan, Benjamin Van Durme, Daniel Khashabi
ManyIH attaches privilege values to individual instructions so models can compare authority directly when directions conflict, rather than relying only on message roles. ManyIH-Bench contains 853 coding and instruction-following tasks with up to 12 privilege levels. Among 13 models, Gemini 3.1 Pro leads at 42.7%, while GPT-5.4 reaches 39.4%; results show benchmark difficulty, not universal agent failure.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.