One Task, Many Interfaces: Simple Cross-Interface Consistency Training Improves LLM Agent Robustness
We propose Cross-Interface Policy Regularization (CIPR), a self-supervised objective that replaces action names with uninformative identifiers and presents each training sample under two interfaces in which equivalent actions are identified by construction.
ProofPaper ↗
Key points
- 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.
- 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.
- We evaluate CIPR in both supervised and reinforcement-learning post-training, where it improves robustness.
Sources (1)
- [1]One Task, Many Interfaces: Simple Cross-Interface Consistency Training Improves LLM Agent RobustnessarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 05:56 PM
We propose Cross-Interface Policy Regularization (CIPR), a self-supervised objective that replaces action names with uninformative identifiers and presents each training sample under two interfaces in which equivalent actions are identified by construction.
Post-training binds the competence of an LLM agent to the interface it was trained on.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 14, 2026$τ$-Emotion-Bench: Exposing the Interaction Fragility of LLM Agents through Real-World User Emotions
- Oct 14, 2026ActionOPD: Distilling Decisions, Not Tokens
- Oct 14, 2026MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments
- Oct 14, 2026Long-Horizon RL Needs Muon
- Oct 14, 2026AgentSafeEvo: Adversarial Co-Evolution for LLM Agent Safety
- Oct 13, 2026Run fast, private agents on-device with LFM2.5-24B-A2B