Hybrid LLM System Excels in Conversational Depression Screening Challenge

The eRisk 2026 Task 1 challenge focused on conversational depression screening, requiring systems to interview LLM personas and infer Beck Depression Inventory II (BDI-II) scores and key symptoms without direct sensitive questions. DS@GT ARC submitted a hybrid multi-agent LLM system that evolved through several stages, culminating in a configuration that leveraged an open-source Gemma 27B model for interviewing. To compensate for the Gemma model's inherent limitations compared to stronger proprietary models, the team integrated three algorithmic components: a precomputed dialogue tree for standardized interview flows, a reliability-weighted consensus aggregation mechanism, and a cluster-based imputation step for unprobed symptoms. This structured algorithmic guidance proved highly effective. The hybrid system achieved an ADODL score of 0.9063, ranking 3rd among all complete submissions and placing DS@GT 2nd overall among 21 teams. Notably, this open-source-driven hybrid configuration outperformed their own paid baseline (GPT-5-nano) while significantly reducing per-persona API costs. This success underscores the potential for algorithmic supervision to enable less powerful, open-source models to achieve competitive results in complex and sensitive conversational AI applications.