One Task, Many Interfaces: Simple Cross-Interface Consistency Training Improves LLM Agent Robustness
Renaming actions cut a trained agent’s accuracy by nearly 25%; the authors report recovering 10.5% without hurting performance on the unchanged interface.
Their method, Cross-Interface Policy Regularization (CIPR), presents each training sample through two interfaces with equivalent actions mapped explicitly. It replaces action names with uninformative identifiers and penalizes differences between the agent’s action distributions. The authors report improved robustness in both supervised and reinforcement-learning post-training. They argue that interface randomization alone still leaves agents vulnerable to interface changes outside the training distribution.
ArXiv · AI/CL/LG's note
Their method, Cross-Interface Policy Regularization (CIPR), presents each training sample through two interfaces with equivalent actions mapped explicitly. It replaces action names with uninformative identifiers and penalizes differences between the agent’s action distributions. The authors report improved robustness in both supervised and reinforcement-learning post-training. They argue that interface randomization alone still leaves agents vulnerable to interface changes outside the training distribution.
ArXiv · AI/CL/LG's note
score 5