AI Paper Atlas

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

ICML 2026 · Large language models › Agents, code and math · approach: Long-horizon LLM agents · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
add explicit safety thoughts block (vs generic think only) (ablation)MOSAIC Qwen2.5-7B · AgentHarm harmfulrefusal rate: 0.59 -> 0.87Improved
add explicit safety thoughts block (ablation)MOSAIC Qwen2.5-7B · ASB DPIcompletion rate: 0.94 -> 0.84Hurt
use pairwise preference reward instead of pointwise scalar (ablation)MOSAIC Qwen2.5-7B · AgentHarm harmfulrefusal rate: 0.79 -> 0.87Improved
use pairwise preference reward instead of pointwise scalar (ablation)MOSAIC Qwen2.5-7B · ASB DPIattack success rate: 0.51 -> 0.42Improved
apply MOSAIC scaffolding to GPT-4oGPT-4o · AgentHarm harmfulharm score: 0.31 -> 0.07Improved
apply MOSAIC scaffolding to GPT-4oGPT-4o · AgentHarm harmfulrefusal rate: 0.00 -> 0.92Improved
MOSAIC training on Qwen2.5-7BQwen2.5-7B · AgentHarm harmfulharm score: 0.18 -> 0.09Improved
MOSAIC training on Qwen3-4B-ThinkingQwen3-4B-Thinking · ASB benigncompletion rate: 0.44 -> 0.85Improved
MOSAIC training on Phi-4Phi-4 · ASB benignbenign refusal rate: 0.43 -> 0.19Improved
MOSAIC training on Phi-4Phi-4 · AgentHarm harmful / ASB DPIrefusal and ASR: refusal 0.94 -> 0.88; DPI ASR 0.19 -> 0.28Hurt
MOSAIC training on Qwen2.5-7BQwen2.5-7B · BFCLv3 base multi-turnaccuracy: 21.0 -> 28.5Improved
MOSAIC training on Qwen2.5-7BQwen2.5-7B · PrivacyLensleakage rate: 0.32 -> 0.26Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy