A new reinforcement learning framework aims to make AI agents less annoying to work with.
Researchers introduced Behavior Agentic Optimization (BAO), a training approach for large language model agents that operate across multiple conversational turns. The core problem it addresses is a real tension: agents that ask too many clarifying questions waste the user's time, but agents that ask too few make mistakes. BAO uses reinforcement learning to regularize what it calls "inter-turn behaviors" — essentially teaching agents when to gather more information and when to just act. Tests against the UserRL benchmark showed BAO outperforming other proactive agent baselines on both task accuracy and reduced user effort.
Most agent research optimizes for one thing: task completion. The tradeoff with user burden rarely gets formal treatment, which is why chatbots still pepper you with three follow-up questions before doing anything. BAO frames this explicitly as a Pareto frontier problem — two competing objectives that can be pushed together rather than traded off — which is a more honest framing than most agent papers offer.
The researchers claim BAO reaches "comparable or even superior" performance to commercial LLM agents, though they stop short of naming which ones. That vagueness is worth flagging: benchmark victories over unnamed commercial systems are a genre of academic claim that deserves a raised eyebrow until the comparisons are made explicit.