Reproduce “LLM Agent Capabilities Should Follow Task Intent and Context Source” 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
Yusheng Zheng, Wenhui Zhang, Yu Mao
IntentCap assembles authority per task, assigning each capability-lease field to its authorized context source. A source-labeling pipeline and LLM-assisted compiler propose short-lived leases, while a deterministic checker verifies provenance and narrowing before enforcement. It had zero unsafe accepts across 3,746 replayed security-sensitive events and covered all 3,813 benign reference actions, limited to tested events and settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.