One Task, Many Interfaces: Simple Cross-Interface Consistency Training Improves LLM Agent Robustness

Artin Tajdini, Manpreet Kaur, Peng Qi|October 14, 2026arXiv

Key Takeaway

Training agents on multiple semantically-equivalent interface variations makes them robust to interface changes without sacrificing performance on the original interface—a simple but effective way to build more generalizable agents.

Summary

This paper addresses how LLM agents become brittle to interface changes (like renaming actions or changing notation) even though the underlying task stays the same. The authors propose Cross-Interface Policy Regularization (CIPR), a training technique that exposes the agent to multiple equivalent interfaces during training, encouraging consistent behavior across them.

training agents

Key Terms

policy-regularization interface-perturbation action-distribution self-supervised-objective