Tool-using agents are typically trained and evaluated against strong simulated users who spell out every requirement explicitly. A new arXiv paper highlights a mismatch: real users rarely behave that way, often leaving out details or expressing themselves incompletely.
The authors propose a simple method to close this gap: construct weak users by masking sentences from a simulated user's replies. This creates training scenarios where the agent must handle incomplete instructions, more closely reflecting real-world interaction.
The paper does not claim that masking fully captures human imperfection, but it offers a concrete step toward agents that are robust to less-than-ideal user input. The approach is notable for its simplicity, though its effectiveness beyond the described setup remains to be seen.