A new arXiv preprint argues that post-training binds an LLM agent's competence to the specific interface it was trained on. This means small, semantics-preserving changes—like renaming an action or switching notation—can degrade performance even though the underlying task remains identical.
The authors propose a simple remedy: cross-interface consistency training. By exposing the model to multiple interfaces that represent the same task, the training encourages the agent to maintain consistent behavior regardless of surface-level differences.
The approach targets a practical problem: real-world systems often update their interfaces, and agents that overfit to one version become brittle. If the method holds up, it could offer a lightweight way to make LLM agents more adaptable without retraining on every possible interface variant.