3.8Zero-Shot Coordination

6 min read

By the end of this section you can answer

Can two agents cooperate on the very first attempt?

Zero-shot coordination evaluates whether an agent can cooperate with an unseen partner without prior joint training, negotiation, or an agreed convention.

In this section you will

• Define zero-shot coordination
• Distinguish zero-shot evaluation from a policy that never adapts
• Explain compatibility in terms of shared conventions
• List the mechanisms that improve first-encounter coordination

Zero-Shot Coordination Definition

Zero-shot coordination

Zero-shot coordination asks whether an agent can cooperate with an unseen partner at deployment, without prior coordination with that partner.

The setup the ZSC literature uses: an ego agent is trained with a limited set of training partners, then evaluated with diverse unseen deployment partners. No joint training, no negotiation, no warm-up round in which the two agree on anything.

“Zero-shot” is about what happened before the encounter. It means no shared history: no co-training, no exchanged conventions, nothing pre-arranged.

Zero-Shot Evaluation and Online Adaptation

Here is the distinction worth getting right.

	What it says	When it happens
Zero-shot coordination	the partner was unseen before deployment	a fact about the setup
Online adaptation	the agent uses interaction during deployment to infer how this partner behaves	a capability of the agent

These are orthogonal, and they compose. An agent can begin an encounter zero-shot, knowing nothing about its partner, and then adapt as evidence arrives, exactly as in Partner Representations. That is still zero-shot coordination: nothing was arranged in advance.

Common pitfall

The misreading to avoid is that zero-shot means the agent may not learn anything about its partner during the episode.

It may. What it may not have is prior coordination. An agent that watches its partner for three steps and then commits is doing zero-shot coordination with online adaptation, and that combination is what most practical systems should aim for.

There is a genuine trade inside this. Adapting takes evidence, and evidence takes steps, and those steps are spent behaving non-committally. On a long task that cost is trivial. On a short one the agent may be better served by a good partner-agnostic prior than by a belief it never has time to sharpen.

Conventions and Compatibility

Zero-shot coordination fails for a specific reason, and the Communicate chapter already produced it.

	Agent A learned	Agent C learned
message 2	bring an ingredient	begin cooking

Both learned successful protocols. Paired zero-shot, 2 means one thing to the sender and another to the receiver, and the team does worse than if neither had spoken.

The Communication Lab measured the general case: six independently trained pairs, 1.00 within pairs, 0.300 across them, some pairings at 0.00.

Key takeaway

Adaptation is not only about physical behaviour. What has to be compatible includes

• roles, who takes which part of the task,
• conventions, which side you pass on,
• communication, what the symbols mean,
• action preferences, what you do when several options are equally good,
• timing, who commits first.

Every one of these is an arbitrary agreement that a co-trained pair settles for free and a zero-shot pair has to get right by luck or by inference.

Mechanisms for Zero-Shot Compatibility

No guarantees here, and it would be dishonest to present these as prescriptions. They are intuitions consistent with the framing, and they are worth holding loosely.

Avoid rigid assumptions. A policy that requires the partner to fetch will fail with any partner that does not. One that checks whether fetching has happened will not.

Respond to observable behaviour. Anything you condition on that the partner actually does is information you genuinely have. Anything you assume is a bet.

Keep more than one interpretation alive. Committing to a partner model after one ambiguous step is how confident-and-wrong happens.

Choose actions that leave room. Where two actions are equally good for you, the one that leaves your partner more options is likely better for the team, because it fails more gracefully if you have misread them.

Prefer interpretable signals. A protocol anchored to something outside the pair has some chance of being read by a stranger, which is exactly the argument of the LangGround connection.

Why these are intuitions

Each of these has counterexamples. Conditioning on observed behaviour is slower than assuming correctly. Leaving room can mean neither agent commits, which is the both-wait failure from Partner Dependence. Interpretable protocols give up expressiveness.

The reason for stating them anyway is that they are the right questions to ask of a design, even when the answers are situational.

Knowledge check

An agent meets an unfamiliar partner, spends four steps observing it, infers its role preference, and then coordinates well. Is this zero-shot coordination?

Select one answer.

AYes. Zero-shot refers to having no prior coordination with that partner. Inferring during the episode is online adaptation, and the two are compatible.BNo. The agent used information about the partner, so it was not zero-shot.CNo, because four steps of observation counts as a warm-up round.DOnly if the agent updates its weights during those four steps.

Check answer

Zero-Shot Coordination Summary

• Zero-shot coordination: cooperating with an unseen partner at deployment, with no prior coordination. An ego agent trained on limited partners, evaluated on diverse unseen ones.
• “Zero-shot” constrains what happened before the encounter, not what the agent may infer during it.
• Online adaptation is a separate axis and composes with it. Beginning zero-shot and then adapting is still zero-shot coordination.
• Adapting costs steps. On short tasks a good partner-agnostic prior may beat a belief there is no time to sharpen.
• Compatibility is needed over roles, conventions, communication, action preferences and timing, all of which a co-trained pair gets for free.
• What tends to help: avoid rigid assumptions, condition on observed behaviour, keep several interpretations alive, leave your partner room, and prefer interpretable signals. These are intuitions, not rules.