Language Game framework enables direct dialogue with non-neural systems

Yanbo Zhang and Michael Levin propose a novel framework, dubbed 'Language Game,' that enables bidirectional dialogue with non-neural systems such as gene regulatory networks, microbial consortia, and fungi. Unlike current approaches where an LLM speaks on the system's behalf, this method lets the system 'speak' in its own voice by treating communication as a Wittgensteinian language game. The system's internal dynamics are frozen as the nonlinear core of a reinforcement-learning policy, while only linear input and output interfaces are trained. Through iterative use and reward, the system's states and responses acquire meaning within the game—playing becomes speaking. A language model acts as a router, selecting the game whose semantics best match a human prompt and designing an environmental state for which the desired action is the rational response, letting the system reply through its own behavior.

Applied across diverse gene regulatory networks and reinforcement-learning tasks, the framework achieves fluent dialogue without altering any system parameters. It demonstrates that well-trained agents from disparate origins converge on similar behaviors, and it reveals that specific properties of gene regulatory networks (e.g., connectivity, nonlinearity) act as an inductive bias, making some systems easier or harder to talk with. This work opens a new path to conversing with any dynamical system on its own terms, with profound implications for synthetic biology, biocomputing, and understanding non-human intelligence.