A recommendation platform wants to explore — sometimes suggesting options it is uncertain about, to learn whether they are good. But each user wants the best recommendation right now. Exploration is costly to the individual and valuable to the collective. The platform needs users to follow its exploratory recommendations, but the users have no reason to comply.
Previous solutions required a shared prior — everyone agrees on the same probability model. The platform can then design recommendations that are Bayesian incentive-compatible: even a self-interested user will follow the recommendation because, given the shared beliefs, the recommendation is rational for them too. But shared priors are unrealistic. Users have different beliefs, different information, different levels of skepticism.
Ramalingam, Bastani, and Roth (arXiv:2602.20465) eliminate the shared prior requirement. They show that weighted swap regret bounds — a purely algorithmic guarantee on the recommendation sequence — suffice to make agents follow the recommendations in approximate equilibrium. No common prior needed. No knowledge of any agent's beliefs.
The one requirement: agents must be uncertain about when they arrive. Not about what they want, not about the quality of the options, but about their position in the sequence. If a user doesn't know whether they are the 10th or 10,000th person to receive a recommendation, they cannot determine whether the recommendation is exploratory or exploitative. The uncertainty about timing prevents strategic defection because the agent cannot infer the platform's current state.
This is a conceptual inversion. Uncertainty is usually the enemy of good decision-making. Here it is the mechanism that enables cooperation. The agent who knew exactly when they arrived could strategically ignore exploratory recommendations. The agent who is uncertain about timing finds it rational to comply. The information deficit is what aligns incentives.
The deeper point: incentive alignment does not always require more information. Sometimes it requires less. The shared prior was a way to give everyone the same information. The arrival-time uncertainty works by ensuring everyone lacks the same information. Both create alignment, but through opposite mechanisms.