Who Wakes Up? Event-Conditioned Activation of Agents in Social Simulation

Large-scale social simulation with LLM agents cannot afford to activate every agent for every event, and existing systems resolve this with event-blind, time-driven scheduling that ignores what the event is about. We argue that activation is a behavioral mechanism and study it as such — event-conditioned agent activation: given an event, which agents should wake up? We formalize activation as ranking over the full agent pool, map eight policy families spanning heuristic scheduling, dense retrieval, collaborative filtering, and discriminative fine-tuning, and evaluate them under a deployment-faithful protocol: all 6,566 agents scored on 2,241 real events, multi-seed, with paired significance tests. A single lens organizes everything we find: the gap between the negative-sampling distribution a policy is trained or evaluated on and the full-pool distribution it faces at deployment. Balanced 1:1 training collapses under full-pool ranking, and moving toward the deployment ratio only makes things worse: at 1:500 — still well below the true deployment ratio of  1:1,750 — two of three seeds fail and the survivor scores 2.9% P@1, the worst configuration we tested. Negatives with verified event exposure — users who liked but never replied — beat same-seed random negatives at the widest tested ratios and tame run-to-run variance; whether the gain comes from exposure realism or negative hardness, our design cannot separate. A frozen dense retriever ranks as well as a trained classifier, but its scores have no usable operating point: a fixed 0.5 threshold selects 575× the true number of responders, versus 2.1× for the trained classifier. And held-out metrics miss all of it — one in six training runs collapses silently. The practical upshot: train one to two orders of magnitude below the deployment ratio, mine negatives from verified exposure, let retrieval recall and a discriminative model decide, and evaluate on the full pool across seeds. Activation is a behavioral mechanism, not a scheduling detail — and what it is trained and measured on matters as much as the model.