Summary
This arXiv paper (2610.20576) by Artin Tajdini, Manpreet Kaur, and Peng Qi introduces Cross-Interface Policy Regularization (CIPR), a self-supervised training objective that improves LLM agent robustness across interface variations. Post-training typically binds an agent's competence to the specific interface it was trained on: semantics-preserving perturbations, such as renaming actions or changing notation, degrade performance without changing the underlying task. The authors show that independent per-sample interface randomization merely broadens this dependence to the training interface distribution, leaving out-of-distribution perturbations harmful. CIPR addresses this by replacing action names with uninformative identifiers and presenting each training sample under two constructed interfaces where equivalent actions are aligned by design, penalizing the KL divergence between the action distributions induced by each interface. Evaluated on both supervised post-training and reinforcement learning post-training, CIPR improves robustness in both settings. On agentic function-calling tasks, held-out interface perturbations reduce post-trained agent accuracy by nearly 25%; CIPR recovers 10.5 percentage points without affecting accuracy on the unchanged interface.
Paper Overview
- Field: NLP
- Authors: Artin Tajdini, Manpreet Kaur, Peng Qi
- Published: 2026-10-14
- arXiv: 2610.20576
Summary
Post-training binds the competence of an LLM 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.
The authors observe that scoring each sample independently means interface randomization only broadens this dependence from a single interface to the training interface distribution, so out-of-distribution perturbations still degrade the agent.
Cross-Interface Policy Regularization (CIPR)
CIPR is a self-supervised objective that:
1. Replaces action names with uninformative identifiers.
2. Presents each training sample under two interfaces in which equivalent actions are identified by construction.
3. Penalizes the KL divergence between the action distributions induced by the two interfaces, forcing the policy to stay consistent across interfaces.
Results
- Evaluated under both supervised post-training and reinforcement learning post-training; CIPR improves robustness in both settings.
- On agentic function-calling tasks, held-out interface perturbations reduce post-trained agent accuracy by nearly 25%.
- CIPR recovers 10.5 percentage points of that accuracy, while leaving accuracy on the unchanged interface unaffected.
---
*Source: arXiv:2610.20576*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178636046