Events

NLP/Text-as-Data Speaker Series: Learning what you know with collaborative self-play

Speaker: Jonathan Berant (Tel-Aviv University, AI2, Google DeepMind)

Location: 60 Fifth Avenue, Room 7th floor open space

Date: Thursday, April 24, 2025

The increasing prevalence of AI assistants requires endowing them with strong collaborative skills. This includes awareness of their parametric knowledge, knowing when to use tools, and adapting to the background and capabilities of their interlocutor. Such capabilities are hard to teach through supervised fine-tuning because they require constructing examples that reflect the agent’s specific capabilities. In this talk, I will describe a new approach for post-training agents to be better collaborators: collaborative self-play. We construct multi-agent collaborations in which the group is rewarded for collectively arriving at correct answers. The desired meta-knowledge emerges from the incentives built into the structure of the interaction. We focus on small societies of agents that have access to heterogeneous tools (corpus-specific retrieval), and therefore must collaborate to maximize their success with minimal effort. Experiments show that group-level rewards for multi-agent communities can improve calibration and tool use and induce policies that transfer to single-agent scenarios. We argue that collaborative self-play is a promising direction for scalable post-training of better AI assistants.