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

TL;DR

The paper tackles miscalibration in visual question answering by introducing AlignVQA, a two-stage multi-agent framework. Specialized VLM-backed agents generate diverse candidate answers, followed by a generalist debate that refines stances and confidences, culminating in a calibrated majority vote. A differentiable AlignCal loss ties training to a provable upper bound on calibration error (UBCE), yielding substantial reductions in ECE, ACE, and MCE across VQARad and ScienceQA. The approach enhances reliability of VQA systems in high-stakes settings, enabling safer and more interpretable autonomous visual reasoning. The work demonstrates that calibrated, debate-driven aggregation plus targeted finetuning yields robust calibration improvements with practical impact.

Abstract

In the context of Visual Question Answering (VQA) and Agentic AI, calibration refers to how closely an AI system's confidence in its answers reflects their actual correctness. This aspect becomes especially important when such systems operate autonomously and must make decisions under visual uncertainty. While modern VQA systems, powered by advanced vision-language models (VLMs), are increasingly used in high-stakes domains like medical diagnostics and autonomous navigation due to their improved accuracy, the reliability of their confidence estimates remains under-examined. Particularly, these systems often produce overconfident responses. To address this, we introduce AlignVQA, a debate-based multi-agent framework, in which diverse specialized VLM -- each following distinct prompting strategies -- generate candidate answers and then engage in two-stage interaction: generalist agents critique, refine and aggregate these proposals. This debate process yields confidence estimates that more accurately reflect the model's true predictive performance. We find that more calibrated specialized agents produce better aligned confidences. Furthermore, we introduce a novel differentiable calibration-aware loss function called aligncal designed to fine-tune the specialized agents by minimizing an upper bound on the calibration error. This objective explicitly improves the fidelity of each agent's confidence estimates. Empirical results across multiple benchmark VQA datasets substantiate the efficacy of our approach, demonstrating substantial reductions in calibration discrepancies. Furthermore, we propose a novel differentiable calibration-aware loss to fine-tune the specialized agents and improve the quality of their individual confidence estimates based on minimising upper bound calibration error.

Figures (6)

• Figure 1: AlignVQA Multi-Agent Calibration Model. Given an input image and question, the model first queries a set of specialized agents—Chain-of-Thoughtwei2022chain, Search-Augmented, SelfAskpress2022measuring, and GENREAD Knowledge-basedyu2022generate models—each fine-tuned for calibration using our custom proposed loss AlignCal . These agents independently produce answer classes (e.g., A: cardinalfish, B: black howler). In the second stage, a group of general agents is instantiated, with each agent probabilistically initialized to a specific answer class based on the distribution of predictions from the specialized agents. These general agents then receive argument-based feedback—comprising for and against justifications—from all general agents (denoted by dotted grey lines), enabling them to revise both their stance and confidence. The final calibrated prediction is the majority-vote class with associated confidence.
• Figure 2: Reliability plots of the datasets, VQARad (top) and ScienceQA (bottom). \ref{['fig:a-gemma']} shows the calibration from base Gemma model. \ref{['fig:b-fl']} shows plot on FL finetuned Gemma 3 4B model. \ref{['fig:c-flalign']} shows the plot on FL + AlignCal finetuned Gemma 3 4B model. \ref{['fig:d-agentic']} shows the plot obtained from Agentic Framework. \ref{['fig:e-agenticalign']} shows the plot obtained from Agentic framework where agents are finetuned with AlignCal + FL.
• Figure 3: Reliability plots for base VLMs on ScienceQA for (a) Gemma‐3‐4B, (b) Qwen2.5‐VL‐3B‐Instruct, (c) LLAVA‐OneVision, (d) Granite‐vision‐3.3‐2B, (e) Phi‐4‐Multimodal‐Instruct, (f) InternVL‐4B, (g) Ristretto‐3B, (h) SmolVLM, (i) Ovis2-4B, and (j) DeepSeekVL2-Tiny. We observe that majority of the base VLM models are overconfident except for \ref{['fig:all_calib_f']}, \ref{['fig:all_calib_g']} and \ref{['fig:all_calib_j']}.
• Figure 4: Reliability plots for base VLMs on VQARad for (a) Qwen2.5-VL-3B-Instruct, (b) Gemma-3-4B, (c) LLAVA-OneVision, (d) Granite-vision-3.3-2B, (e) Ristretto-3B, (f) InternVL-4B, (g) Ovis2-4B, (h) Phi-4-Multimodal-Instruct, (i) SmolVLM, and (j) DeepSeek VL2-Tiny. We observe that majority of the base VLM models are either overconfident or underconfident except for \ref{['fig:rad_f']} and \ref{['fig:rad_g']} on the VQARad dataset.
• Figure 5: Reliability plot comparison between (\ref{['fig:flspath']}) FL and (\ref{['fig:flalignpath']}) FL + AlignCal finetuned LLAVA-OneVision model on VQARad dataset, (\ref{['fig:flscience']}) FL and (\ref{['fig:flalignscience']}) FL + AlignCal finetuned LLAVA-OneVision model on ScienceQA dataset. Our proposed loss FL + AlignCal is effective at improving ECE from 10.5% to 9.3% on the VQARad dataset and a reduction from 9.64% to 6.8% on the ScienceQA dataset. MCE and ACE also follows a similar decreasing trend.