Emergent Communication in Multi-Agent Systems

Language Parameters

Vocabulary size 4

Learning rate 0.15

Training speed 10 ep/s

Latest Round

Waiting for first episode…

Live Stats

Episodes

Success (last 50)

—

Protocol entropy

—

Best so far

How it works

This is a Lewis signaling game — the canonical toy model behind emergent-communication research in multi-agent RL, and the theoretical ancestor of how modern LLM-agent frameworks (AutoGen, CrewAI) coordinate through structured messages instead of a hand-designed protocol.

Each episode: a random target concept is shown to the Sender. It samples a discrete symbol from its own learned distribution over the vocabulary:

P(symbol | concept) = softmax(L_sender[concept])

The symbol crosses the channel with no built-in meaning. The Receiver samples a guess from its own distribution over concepts, conditioned only on the symbol it heard:

P(guess | symbol) = softmax(L_receiver[symbol])

Reward is 1 if the guess matches the target, else 0. Both agents update with REINFORCE (policy-gradient), nudging the logits of the action they actually took toward higher probability when the round succeeded, and away from it when it failed:

advantage = reward − 0.5
L[state][action] += lr · advantage · ( 1[a=action] − P(a) )   for every action a

Neither agent is told what any symbol "means" — a shared code is not designed, it is discovered purely from the shared incentive to succeed together. Watch the protocol entropy readout: it starts near log₂(V) bits (symbols used almost uniformly at random) and falls toward 0 as the sender's mapping collapses onto one confident symbol per concept — the moment a stable language has emerged.

Vocabulary size — fewer symbols than concepts forces the agents to reuse a symbol for more than one object, which caps the achievable success rate.
Learning rate — how aggressively logits move after each round; too high causes the protocol to thrash instead of converging.
Training speed — episodes simulated per second; the 3D scene replays a sampled recent round so you can watch individual transactions even at high speed.

Sender agent Receiver agent Correct round Wrong round

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.