---
title: "Task-Specific Multimodal Question Answering Agents for QANTA 2026 via Confidence Calibration and Incremental Reasoning"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.09623v1"
permalink: "https://feed.astralor.com/items/itm_a350b3a62443409b38f80db0"
published: "2026-07-10T17:22:49.000Z"
score: 77
channel: "finance"
tags: ["GPT-4o","GPT-4o mini","Task-Specific Agent","Multimodal Question Answering","Confidence Calibration"]
---

# Task-Specific Multimodal Question Answering Agents for QANTA 2026 via Confidence Calibration and Incremental Reasoning

## AI commentary

> Although this paper focuses on a single competition, its two-agent architecture is clever: Tossup and Bonus tackle different optimization objectives, and it achieves first place without retrieval or ensembles, relying solely on reasoning strategies and confidence calibration. It offers valuable insights for those working on multimodal QA, agents, or resource-constrained AI systems. I particularly appreciate the dedicated numeric reasoning design to prevent overconfidence, a detail often overlooked in real products.

## At a glance

This paper presents a submission to the QANTA 2026 shared challenge at the ICML 2026 Workshop on Efficient Multimodal Question Answering. Quanta evaluates multimodal quizbowl systems that answer pyramid-style questions from incrementally revealed text and images under realistic efficiency constraints. The challenge involves two tasks: Tossup (deciding when to answer under uncertainty) and Bonus (accurate answer selection and human adoption). The authors develop a task-specific two-agent architecture: the Tossup agent uses a GPT-4o-mini-class model with confidence-calibrated answering and domain-specific numeric reasoning to reduce overconfident predictions; the Bonus agent uses a GPT-4o-class model with leadin-aware reasoning, structured relational reasoning, and multimodal evidence integration for precise answer selection. Without retrieval pipelines or model ensembles, the system achieves the highest overall leaderboard score of 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.

## Key points

- Introduces a task-specific two-agent architecture for Tossup (deciding when to answer under uncertainty) and Bonus (accurate answer selection).
- Tossup agent uses confidence calibration and domain-specific numeric reasoning to reduce overconfident predictions from isolated numeric clues.
- Bonus agent integrates leadin-aware reasoning, structured relational reasoning, and multimodal evidence integration.
- Without retrieval or model ensembles, the system achieves the highest leaderboard score of 0.402 in a hosted-only environment.

## Full text

We present our submission to the QANTA 2026 shared challenge at the ICML 2026 Workshop on Efficient Multimodal Question Answering (EMM-QA). Quanta evaluates multimodal quizbowl systems that answer pyramid-style questions from incrementally revealed text and accompanying images while operating under realistic efficiency constraints. The challenge consists of two distinct tasks: Tossup questions, which require deciding when to answer under uncertainty, and Bonus questions, which emphasize accurate answer selection and human adoption. To address these differing objectives, we develop a task-specific two-agent architecture. Our Tossup agent utilizes a GPT-4o-mini-class model (referred to as GPT-4.1-mini in the competition logs) with confidence-calibrated answering and a domain-specific numeric reasoning policy that reduces overconfident predictions from isolated quantitative clues. Our Bonus agent uses GPT-4o-class model (referred to as GPT-4.1) with leadin-aware reasoning, structured relational reasoning, and multimodal evidence integration to improve exact answer selection. Rather than relying on a retrieval pipeline or model ensembles, our approach emphasizes efficient reasoning policies and confidence calibration within a hosted-only environment. Our system achieved the highest overall leaderboard score of 0.402, including a Tossup score of 0.238 and a Bonus Effect score of 0.164. The results demonstrate that lightweight, task-specific reasoning strategies can provide strong performance on resource-constrained multimodal question answering benchmarks.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_a350b3a62443409b38f80db0