Overview of eRisk 2025: Early Risk Prediction on the Internet (Extended Overview)

Abstract

[Abstract]: This paper presents an extended overview of eRisk 2025, the ninth edition of the CLEF lab on early risk detection. Since its beginnings, eRisk has served as a benchmark for assessing methodologies, evaluation metrics, and challenges in the early identification of personal risks, particularly within health and safety domains. The 2025 edition marks an important evolution, amplifying the lab’s scope toward problems that require richer contextual and conversational understanding. The first task, the only one preserved from last year, asks systems to rank sentences by their relevance to the BDI-II depression symptoms, enabling fine-grained retrieval of depressive cues. The second task reformulates early detection as a contextual decision problem. In this task, the full conversational thread, including the user’s posts and all the interactions from the rest of the people involved, is revealed incrementally. At each step, the models must decide whether sufficient evidence exists to predict depression for the user, thereby rewarding both accuracy and timeliness. Finally, the pilot task pioneers an interactive scenario: fine-tuned large language models engage participants in dialogue and must infer depressive signals from the evolving conversations, probing the feasibility and safety of conversational screening agents. Together, these three tasks continue to advance the field of early risk detection, open new research avenues and align the evaluation framework more closely with real-world conversational settings.