Game theory models learning as a thermostat: agents observe which strategies work, then shift toward better-performing ones with some sensitivity parameter controlling how aggressively they switch. High sensitivity means crisp, rational responses to performance differences. Low sensitivity means noisy, erratic switching that ignores payoff signals. The standard assumption is that sensitivity is fixed — a property of the agent's cognitive apparatus, not a strategic choice. Noise in learning is a limitation to be overcome.
Marta Couto, Fernando Santos, and Christian Hilbe ask what happens when sensitivity itself evolves alongside strategy (arXiv:2506.21498, 2026). They model populations where agents play repeated games and natural selection acts on both the strategy and the learning rule, including how much noise the rule contains.
The results depend on the game. In prisoner's dilemmas, sensitivity increases without bound — evolution pushes toward sharper and sharper responses, approaching perfect rationality. The game rewards precision. But in snowdrift and stag-hunt games, something different happens: sensitivities either converge to a finite value or undergo evolutionary branching, splitting the population into coexisting types with different noise levels.
The finite convergence means evolution finds an optimal amount of imprecision. Too much noise and you ignore useful payoff information. Too little and you respond so aggressively to small differences that you destabilize the equilibrium you're approaching. The branching means the game can sustain a population where some agents are sharp responders and others are noisy ones — and neither type can invade the other completely.
The structural lesson cuts against the rationality assumption. In competitive environments where the best strategy is clear (prisoner's dilemma), cognitive precision is selected for monotonically. But in environments where coordination matters — where the cost of miscoordination exceeds the cost of imprecision — evolution selects for a specific degree of noise. The noise isn't a failure to process information. It's a buffer against the instabilities that perfect information processing creates.
Noise as strategic asset. The trembling hand is not clumsy. It evolved that way.