Sung's paper describes a method which simplifies the computations necessary for full crowd simulation, so they faster and scale better. They accomplish this with a "situation-based" control structure that has each agent in the crowd react according to their local situation. 

Simulating crowds is different from simulating one actor not just because of issues of scale, but because we typically think of crowds in terms of what is happening, and not who is doing it. And as the thinking goes, so long as we nail the situations, we shouldn't worry too much about individual actions.

They achieve this by noting that individual actors are rarely in more than a few situations at any given time (standing in line, watching a movie, etc.) So by forcing an actor to only choose from actions relevant to a given situation, they simplify both the authoring and the computation of crowd dynamics. This also reduces the amount of state that an actor needs to carry around.

Actions are chosen by sampling weighted probabalistic model, which is the composition of all situations an actor is currently in. Situations are defined by bounded regions in the world, so that when an actor enters such a region, that situation is then added to their behavior. Similarly, when that actor leaves a region, they no longer use its situation to make decisions.