Task-Specific Multimodal QA Agents via Confidence Calibration (arXiv:2607.09623)

Summary

This post summarizes a paper by Nirjhar Das and Md. Al-Mamun Provath (arXiv:2607.09623), a submission to the QANTA 2026 shared challenge at the EMM-QA workshop of ICML 2026. The authors developed a task-specific dual-agent architecture for multimodal question answering. The Tossup agent uses a GPT-4o-mini-class model with confidence-calibrated answering and domain-specific numerical reasoning strategies, reducing overconfident predictions on isolated quantitative clues. The Bonus agent uses a GPT-4o-class model with clue-aware reasoning, structured relational reasoning, and multimodal evidence integration. The system achieved the highest total score on the leaderboard at 0.402 (Tossup 0.238, Bonus Effect 0.164), demonstrating that lightweight, task-specific reasoning strategies can deliver strong performance on resource-constrained multimodal QA benchmarks.