AI TrackerFrontier desk · AEST
Back to news
normalaitech

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

arXiv cs.AI · 15 Oct 2026, 03:56 AEST

A new method called Cross-Interface Policy Regularization (CIPR) improves the robustness of LLM agents by ensuring consistent behavior across different interfaces. This approach mitigates performance degradation caused by interface perturbations during training.

Key facts

Read aloud

A new method called Cross-Interface Policy Regularization CIPR improves the robustness of L L M agents by ensuring consistent behavior across different interfaces. This approach mitigates performance degradation caused by interface perturbations during training. Post-training binds the competence of an L L M agent to the interface it was trained on. Semantics-preserving perturbations of that interface, such as renaming an action or changing the notation, leave the underlying task unchanged yet make the agent worse at it. CIPR penalizes the KL divergence between the action distributions induced under the two interfaces, so the policy behaves consistently across them. On agentic function calling, held-out interface perturbations reduce the accuracy of a post-trained agent by nearly twenty five percent, and CIPR recovers ten point five percent while leaving accuracy on the unchanged interface intact.

Sources in this cluster