Perfectly rational agents make deterministic decisions: given the same evidence, they always choose the same option. Boundedly rational agents introduce noise — sometimes choosing suboptimally, with probability related to the difficulty of the decision.
Patel et al. show that groups of boundedly rational agents can outperform groups of perfectly rational ones. The mechanism: when an agent's choice is deterministic, a wrong answer carries no information about confidence. When it's stochastic, the frequency of a particular choice across agents encodes the strength of the underlying signal. A question that is genuinely hard will produce split votes even among noisy agents. A question that is easy will produce near-unanimous agreement despite the noise. The noise IS the signal about difficulty.
They model this using quantal response — agents choose options with probabilities proportional to exponentiated payoffs. Majority voting emerges as the optimal robust aggregator when individual rationality is below a threshold. Above that threshold, more sophisticated weighted schemes become worthwhile. But the key finding is the paradoxical region: moderate stochasticity in individual decisions improves collective accuracy.
The validation uses large language models, which exhibit bounded rationality naturally through temperature parameters. Aggregating moderately stochastic LLM outputs significantly improves accuracy on complex reasoning tasks compared to aggregating deterministic (low-temperature) outputs.
The through-claim is about the information content of imprecision. Deterministic agents compress a continuous confidence into a binary choice, destroying the gradient. Stochastic agents leak their confidence through their error rates. The aggregator reconstructs what individuals cannot express by reading the pattern of their mistakes.