Under what conditions does collective intelligence improve individual as well as societal welfare? Answering this question requires taking seriously complexity of individual and collective behavior, yet most rigorous work—in the service of conceptual and mathematical tractability—prioritizes one level of analysis at the expense of others levels. In this @PNASNews paper, led by @WolframBarfuss, we chart a course towards a principled formal framework for studying how collective cooperative intelligence emerges from the interactions of complex information processing agents learning in group settings. The main value of the paper is that it draws together useful aspects of multiagent reinforcement learning, collective computation and collective behavior, and collective reinforcement learning dynamics, with nods to many other relevant areas of AI and complexity science—like the study of critical points and phase transitions,—to bring into focus the mathematical properties a principled framework needs to have to move beyond toy model insights and build a theory of collective intelligence that can inform intervention in real world systems, as well as help build better artificial intelligence. pnas.org/doi/pdf/10.1073/pna…