Autonomous agents and robots are taking on increasingly collaborative roles alongside people, from self-driving vehicles sharing the road with human drivers to language-model agents executing tasks on a user's behalf. In each setting, success depends not only on individual task competence, but on the ability to work effectively with partners whose preferences and strategies shape the outcome. This thesis explores collaborative intelligence, focusing on the components required to move from static, purely reactive agents to proactive collaborators that reason, adapt to, and shape the behaviors of their teammates.
We first evaluate collaborative teamwork through ChefBench, a benchmark for human-agent coordination and ad hoc teaming built on an extended version of Overcooked with richer coordination challenges and a real-time, browser-based interface. Human studies and state-of-the-art MARL baselines show that ChefBench elicits far greater strategic diversity than prior Overcooked variants and exposes a substantial, persistent performance gap between human and agent teams, alongside a released dataset of 147 human-human gameplay trajectories.
We then address adaptation with TALENTS, an ad hoc teamwork algorithm that infers a teammate's strategy from a learned latent behavior space and adapts its policy accordingly. TALENTS clusters offline trajectories into discrete teammate types and uses a regret-minimization algorithm to track and respond to a partner's strategy as it evolves within an episode. Across agent-agent evaluations and a 119-participant human-agent study, TALENTS outperforms existing baselines in both task reward and subjective measures of team fluency and trust.
Finally, we extend beyond adaptation to proactive influence, examining how an agent that models how its teammate learns can deliberately shape that learning process to steer partners toward more effective joint conventions, rather than simply best-responding to a fixed strategy.
Together, these contributions establish algorithmic foundations for autonomous agents that not only adapt intelligently to human and artificial teammates, but actively help shape more effective collaboration.