One Task, Many Interfaces: Simple Cross-Interface Consistency Training Improves LLM Agent Robustness
CIPR trains LLM agents to stay consistent across renamed or notationally changed interfaces.
The paper argues that post-training binds an LLM agent to the interface it saw, so semantics-preserving changes such as renamed actions or altered notation degrade performance. Cross-Interface Policy Regularization presents each sample under two equivalent interfaces and penalizes KL divergence between their action distributions. On agentic function calling, held-out interface perturbations reduce accuracy by nearly 25%, and CIPR recovers 10.5 percentage points while leaving accuracy on the unchanged interface intact.
- Interface changes such as renamed actions degrade trained LLM agents.
- CIPR pairs two interfaces and penalizes KL divergence between action distributions.
- Held-out perturbations cut function-calling accuracy by nearly 25%.
- CIPR recovers 10.5 points without hurting the original interface.
- The objective works in both supervised and reinforcement-learning post-training.
From the original article
1 min read at arxiv.orgPost-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.
By Artin Tajdini · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.