Update README.md

README.md

@@ -23,6 +23,7 @@ pipeline_tag: text-classification
 It identifies whether an agent’s long-horizon trajectory complies with externally imposed or human-specified policies.  
 Built on the **PolicyGuardBench** dataset, it achieves **high accuracy**, **cross-domain generalization**, and **remarkable efficiency** — proving that compact guardrails can be both **accurate** and **deployable**.
 
+For more details, refer to our paper: *[Towards Policy-Compliant Agents: Learning Efficient Guardrails For Policy Violation Detection](https://arxiv.org/abs/2510.03485)*.
 ---
 
 ## 🧩 Key Objectives