Abstract:
Misinformation governance must also address cases in which false narratives have already spread and elicited heterogeneous group responses. We formulate this setting as cognitive-impact mitigation on observable propagation graphs and propose a graph intervention agent with decoupled structural and decision layers. The structural layer models propagation graphs, estimates node states, constructs group candidates, and performs standardized simulation replay. The decision layer uses a large language model agent to compare candidates, audit corrective messages, and submit actions within a controlled tool space. Experiments cover
13660 propagation samples from CED, PHEME, Weibo21, and MisBot with three language models. In the main experiments, methods are compared on given propagation-graph snapshots under common per-sample budget caps and a unified simulation replay protocol. The proposed method achieves a higher mean simCHR than the controls reported in these experiments, and its paired improvement over the one-shot baseline is positive in all 12 dataset–model combinations. Additional analyses show that the group candidate space and state-aware scoring are major sources of replay gain. With intervention targets fixed, agent-generated corrective messages receive higher expert-rated expected belief-reduction scores than generic templates. These results show that the proposed agent achieves positive proxy harm-mitigation gains in simulated post-propagation group intervention.