Tracing the 2024 U.S. election debate on Telegram with LLMs and graph analysis

We examine conversations on Telegram during the 2024 U.S. elections to understand how political narratives emerge and cluster at scale. We propose a general-purpose pipeline that combines message-level topic modeling with co-forwarding graph analysis to filter thematically relevant chats. LLM-based daily summarization and encoding are then applied to detect topics and trace the dynamics of chat attention over time in large-scale conversational datasets. Applied to 486 M messages, our method isolates politically engaged groups and detects 36 refined topics active during June–July 2024. We uncover cohesive thematic spheres-clusters of chats with synchronized attention and selective content sharing-that include ideologically extreme or conspiratorial niches. The framework generalizes beyond this case, providing a scalable tool for studying narrative alignment in messaging platforms and social networks.