Two strangers hit it off in the conference lunch line — then reach the food and face the real problem: how do you part without either of you having to admit you'd like to? It's a small, universal agony, and it's a proper game-theory problem. Here are six models of it, each with knobs you can turn.
A companion to the conference-room simulation — that one watches a whole room self-organize; this one zooms in on the single most awkward decision inside it. The models are real; the words are a first draft to argue with.
The setup: two people meet in line, have a genuinely good conversation, and then must navigate whether to keep going together after they've got their food. The trouble is that neither knows what the other actually wants, saying what you want carries a social cost, and — at a multi-day conference — you might run into them again. That's incomplete information, costly signalling, and a repeated game, all inside ninety seconds of small talk.
Strip out the uncertainty first. Two players, each chooses Stay or Part. Each has a true type: Type W ("wants to continue") gets utility from mutual continuation, and Type E ("wants to escape") gets utility from parting. For a Type W player the payoffs rank a > d > b ≈ c — mutual staying is best, a clean mutual parting is fine, and the mismatches (lingering alone, or leaving with a pang of guilt) are worst. For Type E the ranking flips to d > a > c > b: getting out cleanly is the prize, and being trapped in unwanted continuation is the nightmare.
Type E's matrix lives at the same four cells but with the ranking d > a > c > b: a clean mutual exit (d) is the prize, mutual continuation (a) is merely tolerable, escaping while they stay (c) beats the nightmare of staying while they leave (b). (A note for the formally-minded: I take the worst outcome to be b, per that ordering — it's what makes the "two pure equilibria" below actually hold.)
Now put a W and an E together and something clean falls out. Both still prefer to match — coordinating beats a mismatch for either of them — but they root for opposite equilibria: W for (Stay, Stay), E for (Part, Part). That's a Battle of the Sexes. It has exactly the two pure equilibria the model names — plus a mixed one — and, crucially, the payoffs alone can't pick the winner. Only beliefs about the other's type can, which is precisely what pushes us into the incomplete-information world of the next models.
The knobs below let you set either type's payoffs — locked to that type's ranking, because the two equilibria are an ordinal fact you can't slide away. What the sliders genuinely move is the mixed equilibrium: how likely the two of you are to miscoordinate. Flip the W/E toggle to see each matrix, and who wants which exit:
Now neither player sees the other's type. Let p be the prior probability that any given person is Type W, so 1 − p is the chance they're Type E. A strategy is no longer a single action but a rule mapping your type to an action, which gives four pure profiles:
The separating profile is the efficient one: everyone's action broadcasts their type and coordination just works. But it only survives if parting carries no social cost — and in the real world it always does. So it collapses, and what's left standing is pooling on Stay: the equilibrium we actually observe, which holds precisely when the social cost of signalling "I'd like to leave" outweighs the cost of unwanted continuation. The next model makes that cost explicit — and lets you watch it erase information.
Before choosing, each player can send a message m ∈ {Keen, Neutral, Exit}. The catch is a cost asymmetry under social pressure: sending "Keen" is free for a Type W but costs a Type E a penalty s > 0 (the discomfort of performing enthusiasm you don't feel); sending "Exit" is free for E but costs W the same s (you've signalled disinterest and risked rejection). That s is the social-grace penalty, and it's the villain of the whole story.
The receiver updates on the message by Bayes' rule:
Here's the sting. In the pooling equilibrium both types send "Keen", so Pr(Keen | W) = Pr(Keen | E) = 1, and the whole expression collapses:
That is the awkwardness, formalized. The very message that's supposed to help you coordinate tells the listener nothing, because everyone sends it. Turn the knobs below: drag the two "sends Keen" probabilities together and the posterior snaps back to the prior (pooling, no information); pull them apart and "Keen" starts to mean something again (separating).
Why does pooling persist if it's so useless? Formally it's fragile — a Type E always has a private incentive to peel off toward an honest "Exit":
For Type E, u(P, P) > u(P, S), so the right side wins and the inequality fails for any s > 0. The reason pooling is nonetheless what we see is that s is endogenously large — among short-lived strangers with no reputation on the line, the felt cost of looking rude swamps the coordination benefit.
Enter the master move: the escape-route offer — "I'm sure you want to go find your people…". It's a message that does three things at once: a partial signal that you'd part, a face-saving gift to the other person, and something plausibly deniable as mere politeness. Formally it lowers the receiver's cost of parting by δ, the value of having been given permission:
That extra δ shifts the receiver's threshold, coaxing out more honest responses and partly rebuilding the separating equilibrium that grace had knocked down. There's a catch — the sender's risk: if a Type W floats the offer as pure politeness and a Type E gratefully accepts, the sender lands in their worst outcome (they stay, the other leaves). So in equilibrium the offer is sent more often by people who actually want out — which is exactly what makes it a partially informative signal. Turn up δ and watch honest parting recover:
If the players move in sequence rather than at once, the first mover leaks a little type information and the second can respond to it — coordination improves. But the same pooling problem caps the gain: if both types would send the same opener, the second mover still learns nothing. The advantage only materializes when the first signal is credible, i.e. genuinely costly to send. Three real-world devices manufacture that credibility:
A one-shot encounter only weighs today's payoff. A multi-day conference adds T rounds of possible re-encounter, and with them, reputation. Let r be the other person's belief that you're warm and worth knowing; a Type E's value becomes their immediate payoff plus a discounted stream of future reputation:
with f(r) increasing in reputation. Because that sum grows with the horizon, even someone who wants out now has a reason to invest in a graceful exit — to manage the departure so r survives it.
This is why conference goodbyes are stuffed with future-tense promises — "let's grab coffee later," "I'll email you" — that both people know won't happen. They aren't coordination commitments; they're reputation-preserving moves in the repeated game. Turn up the horizon and watch the empty promise turn from a needless cost into a paying investment:
The through-line across all six models is a single, slightly bleak result: social grace actively destroys the information needed to coordinate. The politeness that makes the encounter pleasant is exactly what makes the exit hard — it prices honest signals out of the market, leaving everyone pooled on a cheerful, uninformative "Keen." The escape-route offer is the closest thing to a norm-compatible way back out: a signal you're allowed to send.
Open threads I'd still like to pull on: the explicit mixed-strategy equilibria, reputation r(t) as a stochastic process, a loss-aversion angle on face-saving, and the empirical question underneath all of it — what is the share of Type W in a given room, and how does it move with the kind of conference you're at?
Taking a fuzzy human situation, writing down the incentives, and turning it into something you can actually turn the dials on — that's the fun part of the job.