Guozhang Chen, Pulin GongSelective attention can sift out particular objects or features from the plethora of stimuli. Such preferential processing of attention is often compared to a spotlight pausing to illuminate relevant targets in visual fields in a stimulus-driven way (bottom-up attention) and/or task-driven way (top-down attention). Recent studies have revealed that bottom-up distributed attention involving multiple objects is not a sustained spotlight, but samples the visual environment in a fundamentally dynamical manner with theta-rhythmic cycles, with each sampling cycle being implemented through gamma oscillations. However, the fundamental questions regarding the dynamical nature and the circuit mechanism underlying such dynamical attentional sampling remain largely unknown. To address these questions, in this study we investigate a biophysically plausible cortical circuit model of spiking neurons and find that in the working regime of the model (i.e. the regime near the critical transition between the asynchronous and propagating wave states), the localized activity pattern emerging from the circuit exhibits rich spatiotemporal dynamics. We illustrate that the nonequilibrium nature of the localized pattern enables the circuit to dynamically shift to different salient external inputs, without introducing additional neural mechanisms such as inhibition of return as in the conventional winner-take-all models of attention. We elucidate that the dynamical shifting process of the activity pattern provides a mechanistic account of key neurophysiological and behavioral findings on attention, including theta oscillations, theta-gamma phase-amplitude coupling, and vigorous-faint spiking fluctuations. Furthermore, by using the saliency maps of natural stimuli, we demonstrate that the nonequilibrium activity pattern dynamics can better explain the psychophysical findings regarding attention maps and attention sampling paths than the conventional models, providing a profound computational advantage for efficiently sampling external environments. Our work thus establishes a novel circuit mechanism by which non-equilibrium, fluctuating pattern dynamics near the critical transition of circuit states can be exploited for implementing efficient attentional sampling.