Learning When to Act or Refuse: Guarding Agentic Reasoning Models for Safe Multi-Step Tool Use

Researchers have introduced two new frameworks, MOSAIC and AutoTool, to improve the safety and functionality of agentic language models used for multi-step tool use and reasoning.

MOSAIC is a post-training framework that aligns agentic language models for safe multi-step tool use by structuring inference as a plan, check, then act or refuse loop with explicit safety reasoning^[1]. It uses preference-based reinforcement learning with pairwise trajectory comparisons to train without trajectory-level labels. MOSAIC has been shown to reduce harmful behavior by up to 50% and increase harmful-task refusal by over 20% on injection attacks^[1]. AutoTool, on the other hand, is a training framework that enables dynamic tool selection and integration for agentic reasoning through a dual-phase optimization pipeline^[2]. AutoTool has outperformed advanced LLM agents and tool-integration methods on ten diverse benchmarks, achieving a 6.4% average gain in math & science reasoning, 4.5% in search-based QA, 7.7% in code generation, and 6.9% in multimodal understanding^[2].