Reproduce “Learning When to Act or Refuse: Guarding Agentic Reasoning Models for Safe Multi-Step Tool Use” 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
Aradhye Agarwal, Gurdit Siyan, Yash Pandya, Joykirat Singh, Akshay Nambi, Ahmed Awadallah
MOSAIC trains agents to make safety checks and refusals explicit choices within their action loop. After planning, an agent may assess risks before acting, answering, asking for clarification, or taking a terminal refusal action; pairwise preferences over complete trajectories guide training. On Qwen2.5-7B, AgentHarm harm falls from 0.18 to 0.09, but gains vary by model and metric.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.