Researchers have developed a novel two-agent architecture for the QANTA 2026 challenge, designed to excel in multimodal question answering under efficiency constraints. The system employs a GPT-4o-mini-class model for Tossup questions, incorporating confidence calibration and a numeric reasoning policy to mitigate overconfidence. A separate GPT-4o-class model handles Bonus questions, utilizing lead-in awareness, relational reasoning, and multimodal evidence integration for precise answer selection. This approach achieved the highest overall score of 0.402 on the leaderboard, demonstrating the effectiveness of task-specific reasoning and calibration in resource-constrained environments. AI
IMPACT Demonstrates effective strategies for multimodal Q&A under efficiency constraints, potentially influencing future system design.
RANK_REASON The cluster contains a research paper detailing a novel AI system submitted to a challenge.
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
