Hybrid LLM system ranks 3rd in depression screening challenge

Researchers from DS@GT have developed a hybrid multi-agent LLM system for conversational depression screening, achieving a 3rd place ranking in the eRisk 2026 Task 1 challenge. Their system, which interviews LLM personas simulating individuals with depression, evolved from a single-model prototype to a multi-agent architecture and finally to a hybrid configuration. This hybrid approach replaced a proprietary GPT-5 nano interviewer with the open-source Gemma 27B model, enhanced by algorithmic components like a precomputed dialogue tree and a reliability-weighted consensus aggregation. The hybrid system demonstrated competitive performance, outperforming the proprietary model at a significantly lower cost. AI

IMPACT Demonstrates the effectiveness of algorithmic guidance in enabling open-source models to compete with proprietary ones for specialized tasks.

RANK_REASON Research paper detailing a novel hybrid multi-agent LLM system for a specific task. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• DagsHub
• DS@GT
• eRisk 2026
• Gemma 27B
• GPT-5 nano
• Hugging Face
• Weaver framework

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →