Task-specific agents with confidence calibration win QANTA 2026

The authors present a submission to the QANTA 2026 shared challenge at the ICML 2026 Workshop on Efficient Multimodal Question Answering (EMM-QA), achieving the highest overall leaderboard score of 0.402.

• The system uses a two-agent architecture tailored to Tossup and Bonus questions.
• The Tossup agent employs GPT-4o-mini-class with confidence-calibrated answering and numeric reasoning to reduce overconfident predictions.
• The Bonus agent utilizes GPT-4o-class with leadin-aware reasoning and multimodal evidence integration for accurate answer selection.
• The approach avoids retrieval pipelines or model ensembles, relying instead on efficient reasoning policies within a hosted-only environment.

The results demonstrate that lightweight, task-specific reasoning strategies can provide strong performance on resource-constrained multimodal question answering benchmarks.