Emergent Communication — Policy Graph (2D)

Two independent agents — a Sender that sees a target object and a Receiver that only hears a symbol — are trained end-to-end with nothing but a shared reward for guessing correctly. This 2D counterpart renders the resulting Lewis signaling game as a live bipartite policy graph: five concept icons on the left, the current vocabulary of symbol nodes in the middle, and the same five concepts mirrored on the right as the receiver's possible guesses. Every edge's thickness and opacity is drawn directly from the agents' softmax probability tables, so the entire learned language is visible at once, and a traveling token replays a sampled round across it. Rolling strip charts beneath the graph track success rate and protocol entropy across training — a distinct, graph-and-timeseries representation of the exact same REINFORCE mechanics used by the 3D version, not a flattened camera view of it.