Research map / Multi-agent systems and games / LLM agents

Refine and Align: Confidence Calibration Through Multi-Agent Interaction in VQA

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AlignVQA, a multi-agent debate framework over four specialized VLMs plus a differentiable calibration loss (AlignCal) to reduce overconfidence in multiple-choice VQA, evaluated on ScienceQA and VQA-RAD.

Problem

Vision-language models used for VQA are often overconfident, so their confidence scores do not reflect actual correctness, which is risky in high-stakes settings such as medical diagnosis.

Approach

Combine a two-stage debate among specialized VLM agents (different prompting strategies) and generalist agents that critique and revise stances, with fine-tuning via a differentiable surrogate of the upper-bound calibration error (AlignCal) added to focal loss.

Key results

• Agentic debate reduces ECE on VQA-RAD from 0.375 (Gemma 3 4B base) to 0.146 and ACE from 0.207 to 0.133 (as reported in the intro).
• AlignCal fine-tuning with focal loss reduces ECE on ScienceQA from 0.232 to 0.058 for Gemma 3 4B (intro/results), and the best table entry reports ECE 0.055 for Agentic+AlignCal+FL on ScienceQA.
• Ablation: increasing debate rounds hurts calibration (ECE 0.146 to 0.188), so a single round is used.

Surprising findings

• Increasing the number of debate rounds worsened calibration rather than improving it.
• Dropping weaker agents (4 to 3) improved calibration beyond any individual stage-1 agent.

Limitations

• Inference cost is high due to two sequential debate stages across multiple agents.
• VQA-RAD evaluated only on Yes/No questions and the paper reports inconsistent numbers across sections (e.g., ECE values differ between text, tables and figures).
• MCE is not guaranteed by the loss, as it minimizes an expectation; authors acknowledge only practical improvements.

Experimental setup

Datasets: ScienceQA VQA-RAD

Benchmarks: ScienceQA VQA-RAD

Metrics: accuracy F1 precision recall ECE ACE MCE

Baselines: Focal Loss fine-tuning Label Smoothing Temperature Scaling Dirichlet Calibration LLaVA-OneVision Gemma 3 4B Qwen2.5-VL-3B-Instruct Phi-4-multimodal-instruct

Builds on: Collaborative Calibration focal loss temperature scaling Dirichlet calibration label smoothing chain-of-verification

Closest papers by meaning

• Mind the Uncertainty in Human Disagreement: Evaluating Discrepancies Between Model Predictions and Human Responses in VQA (AAAI 2025 · similarity 0.89)
• CertainlyUncertain: A Benchmark and Metric for Multimodal Epistemic and Aleatoric Awareness (ICLR 2025 · similarity 0.88)
• QA-Calibration of Language Model Confidence Scores (ICLR 2025 · similarity 0.88)
• FRAMES-VQA: Benchmarking Fine-Tuning Robustness across Multi-Modal Shifts in Visual Question Answering (CVPR 2025 · similarity 0.87)
• Automated Generation of Challenging Multiple-Choice Questions for Vision Language Model Evaluation (CVPR 2025 · similarity 0.87)
• Improving Automatic VQA Evaluation Using Large Language Models (AAAI 2024 · similarity 0.87)