Researchers have developed a novel two-agent architecture for multimodal question answering, specifically designed for the QANTA 2026 challenge at the ICML 2026 Workshop on Efficient Multimodal Question Answering. The system employs a confidence-calibrated agent using a GPT-4o-mini-class model for deciding when to answer under uncertainty, and a GPT-4o-class model for accurate answer selection and human adoption. This approach, emphasizing efficient reasoning and confidence calibration over ensembles, achieved the highest overall score of 0.402 on the QANTA leaderboard. AI
IMPACT Demonstrates effective task-specific reasoning strategies for resource-constrained multimodal QA systems.
RANK_REASON This is a research paper submission to a specific academic challenge. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
