Research map / Large language models / Agents, code and math
Learning When to Act or Refuse: Guarding Agentic Reasoning Models for Safe Multi-Step Tool Use
Aradhye Agarwal, Gurdit Siyan, Yash Pandya, Joykirat Singh, Akshay Nambi, Ahmed Hassan Awadallah
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add explicit safety thoughts block (vs generic think only) (ablation) | MOSAIC Qwen2.5-7B · AgentHarm harmful | refusal rate: 0.59 -> 0.87 | Improved |
| add explicit safety thoughts block (ablation) | MOSAIC Qwen2.5-7B · ASB DPI | completion rate: 0.94 -> 0.84 | Hurt |
| use pairwise preference reward instead of pointwise scalar (ablation) | MOSAIC Qwen2.5-7B · AgentHarm harmful | refusal rate: 0.79 -> 0.87 | Improved |
| use pairwise preference reward instead of pointwise scalar (ablation) | MOSAIC Qwen2.5-7B · ASB DPI | attack success rate: 0.51 -> 0.42 | Improved |
| apply MOSAIC scaffolding to GPT-4o | GPT-4o · AgentHarm harmful | harm score: 0.31 -> 0.07 | Improved |
| apply MOSAIC scaffolding to GPT-4o | GPT-4o · AgentHarm harmful | refusal rate: 0.00 -> 0.92 | Improved |
| MOSAIC training on Qwen2.5-7B | Qwen2.5-7B · AgentHarm harmful | harm score: 0.18 -> 0.09 | Improved |
| MOSAIC training on Qwen3-4B-Thinking | Qwen3-4B-Thinking · ASB benign | completion rate: 0.44 -> 0.85 | Improved |
| MOSAIC training on Phi-4 | Phi-4 · ASB benign | benign refusal rate: 0.43 -> 0.19 | Improved |
| MOSAIC training on Phi-4 | Phi-4 · AgentHarm harmful / ASB DPI | refusal and ASR: refusal 0.94 -> 0.88; DPI ASR 0.19 -> 0.28 | Hurt |
| MOSAIC training on Qwen2.5-7B | Qwen2.5-7B · BFCLv3 base multi-turn | accuracy: 21.0 -> 28.5 | Improved |
| MOSAIC training on Qwen2.5-7B | Qwen2.5-7B · PrivacyLens | leakage rate: 0.32 -> 0.26 | Improved |
Closest papers by meaning
- AlphaAlign: Incentivizing Safety Alignment with Extremely Simplified Reinforcement Learning (ICLR 2026 · similarity 0.91)
- The Alignment Waltz: Jointly Training Agents to Collaborate for Safety (ICLR 2026 · similarity 0.89)
- Think Twice Before You Act: Enhancing Agent Behavioral Safety with Thought Correction (ICML 2026 · similarity 0.89)
- MLLMs Fail to Refuse when Using Tools Agentically (NeurIPS 2026 · similarity 0.89)
- Contrastive Reasoning Alignment: Reinforcement Learning from Hidden Representations (ICML 2026 · similarity 0.89)
- Internalizing Safety Understanding in Large Reasoning Models via Verification (ICML 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
