ALIGN: Agent‑Environment Alignment via Automated Interface Generation

---

🔍 What is ALIGN?

Large‑language‑model (LLM) agents often fail because their interfaces (the action & observation layer that sits between the agent and the environment) neglect to reveal hidden pre‑conditions or constraints. ALIGN is a system that automatically generates a richer, better‑aligned interface without touching either the agent logic or the environment code.

Key Contributions

Our key contributions can be summarized as follows:

• We identify and characterize the agent-environment misalignmeent problem, empirically demonstrating its prevalence across diverse domains and its role as a siggnificant bottleneck to agent performance.
• We introduce ALIGN, the first framework automatically generates alligned interfaces to alleviate agent-environment misalignment, without modifying agent logic orenvironment code.
• We demonstrate the effectiveness and generalizability of ALIGN across three domains, with up to a 45.67% success rate improvement on ALFWorld and consistent boosts on ScienceWorld, WebShop, and M3ToolEval.

---

✨ Results

---

✏️ Citing ALIGN

If you find this work useful, please cite:

@misc{liu2025agentenvironmentalignmentautomatedinterface,
      title={Agent-Environment Alignment via Automated Interface Generation}, 
      author={Kaiming Liu and Xuanyu Lei and Ziyue Wang and Peng Li and Yang Liu},
      year={2025},
      eprint={2505.21055},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2505.21055}, 
}

---

💬 Contact & Questions

Open an issue or reach us at [email protected].

Enjoy and happy aligning! 🎉