Attention has a metabolic price tag, and a new study says the smartest way to spend it depends on the stakes.
Researchers built a normative model of a signal-detection task, the kind where an agent must notice a brief, sometimes rare signal, and calculated what allocation of costly attention would perform best. The model found two distinct optimal strategies. When the cost of intense focus is steep, the best approach is to ramp attention up gradually as evidence accumulates. When focus is cheaper, attention instead fluctuates rhythmically, pulsing on and off rather than holding steady, and that pulse frequency climbs with the reward for a correct detection and with how brief or frequent the target signal is.
The result gives mathematical backing to something attention researchers have observed for years but struggled to explain: human attention during long vigilance tasks often oscillates instead of staying locked on. The model traces rhythmic attention to agents that lean on a stable internal sense of timing rather than constantly updating based on new sensory evidence, a distinction that matters for anyone designing AI systems that must budget compute or sensor time instead of running at full attention around the clock.
It is a reminder that constant, maximal attention is not the gold standard. Sometimes blinking is the efficient move.