Who Acts, to Whom, and What They Say: Action-Aware Social Simulation at Scale

Behavioural foundation models answer what a given person would say: the speaker is an input. A population simulator must also decide who acts, when, and to whom, and today these decisions are typically left to hand-written rules. We pose full-event simulation, a task in which the speaker is an output: from a platform's history, a model emits the next events (how many, when, who, to whom, and what they say) and continues from its own output. We build ACT-AWARE, which learns who acts and whom they answer, and writes what they say with a post-trained behavioural foundation model. On a single-step version of the task, ACT-AWARE names real actors with 49.6% precision, against 36.6% for the best of fifteen language models prompted zero-shot. Rolled out the natural way, such a simulator meets a limit no model can lift. If it draws actors from a trailing window of recent posters and feeds on its own output, its coverage, the share of real actors it selects, has a ceiling fixed before its first step. A controller lets ACT-AWARE widen that window to seven days with no observed drop in precision. Over five held-out days, its coverage reaches 2.35× the ceiling of a six-hour window, against 1.27× for random ranking and 1.85× for ranking by lifetime post count, both under the same controller and seven-day pool. Structurally, its output also becomes the world later events come from: 82.4% of its simulated replies attach to posts it generated, against 85.2% in real data. In short, who acts is a learnable factor rather than a scheduling convention.