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

AI summary

Overview

• Research area: Confidence calibration for vision-language models (VLMs) in multiple-choice Visual Question Answering (VQA), using a multi-agent debate framework combined with a differentiable calibration-aware training loss.
• Technical level: Advanced. The paper combines agentic multi-agent orchestration, prompt-taxonomy engineering, and a statistical-learning derivation of a plug-in surrogate for an upper bound on calibration error.
• Scope: One sentence — the paper proposes AlignVQA, a two-stage debate-and-refinement pipeline over specialized and generalist VLM agents, plus AlignCal, a calibration-aware loss, and measures the resulting reductions in ECE, ACE and MCE on ScienceQA and VQA-RAD.

What This Paper Is About

Modern VQA systems built on VLMs have become more accurate, but their reported confidence often does not match how often they are actually correct — they tend to be overconfident, including when they are wrong. In high-stakes settings such as medical imaging or autonomous navigation, an overconfident incorrect answer is a safety problem, not just a quality problem. The paper's goal is to produce VQA systems whose confidence better reflects true correctness, using (i) structured debate among multiple agents and (ii) a training loss that directly penalizes miscalibration rather than only penalizing wrong answers.

Key Contributions

1. AlignVQA, a multi-agent debate framework for post-hoc VQA calibration. Specialized agents with different VLM backbones and different prompting strategies generate candidate answers; generalist agents then critique, revise and aggregate them across two stages. The final answer is chosen by majority vote over the refined stances, and the final confidence is the mean refined confidence of the agents supporting that stance.
2. AlignCal, a differentiable calibration-aware loss. The loss is a plug-in surrogate for the Upper Bound Calibration Error (UBCE) and is minimized jointly with focal loss, giving the combined objective L_tot = L_FL + λ·L_AlignCal. The paper derives the surrogate, analyzes its gradients, and argues that minimizing it directly tightens a provable upper bound on calibration error.
3. Fine-tuning specialized agents for calibration. Each specialized agent is fine-tuned with AlignCal, and the paper reports that better-calibrated specialized agents yield better-aligned confidences inside the debate.
4. Empirical evaluation across two benchmarks. Experiments on ScienceQA and VQA-RAD compare the agentic framework and AlignCal against base VLMs, focal loss, label smoothing, temperature scaling and Dirichlet calibration.

Main Findings

• Base VLMs are substantially miscalibrated. In Table 1, ScienceQA ECE ranges from 0.302 (Qwen2.5-VL-3B-Inst.) to 0.574 (Phi-4-multimodal-Inst.), and VQA-RAD ECE ranges from 0.134 (Phi-4-multimodal-Inst.) to 0.375 (Gemma 3 4B), with Gemma 3 4B showing MCE of 0.818 on VQA-RAD.
• Debate alone improves calibration. With the agentic framework, the paper reports ScienceQA ECE falling from 0.396 to 0.270, ACE from 0.372 to 0.265 and MCE from 0.449 to 0.438; on VQA-RAD, ECE falls from 0.375 to 0.143 and ACE from 0.207 to 0.144.
• AlignCal alone improves calibration more sharply on ScienceQA. The paper reports ScienceQA ECE reduced from 0.232 to 0.058 and ACE from 0.232 to 0.120 with AlignCal, and on VQA-RAD ECE from 0.178 to 0.137 and ACE from 0.155 to 0.115.
• Debate plus calibrated agents gives the largest combined effect. With agents fine-tuned using AlignCal, the debate yields VQA-RAD ECE 0.375 → 0.098, MCE 0.818 → 0.267 and ACE 0.207 → 0.095; on ScienceQA, ECE 0.396 → 0.055, ACE 0.372 → 0.110 and MCE 0.449 → 0.331.
• The proposed method is competitive on accuracy while leading on calibration (Table 2). Agentic + AlignCal + FL reaches 68.20% accuracy, ACE 0.095, ECE 0.098 and MCE 0.267 on VQA-RAD, and 76.10% accuracy, ECE 0.055, ACE 0.110 and MCE 0.331 on ScienceQA.
• Training-time baselines are weaker than the proposed loss. With Agentic + FL, VQA-RAD shows ECE 0.073 and MCE 0.393, while ScienceQA shows ECE 0.180 and MCE 0.678; with Agentic + LS, VQA-RAD shows ECE 0.183 and ScienceQA ECE 0.186 with MCE 0.916.
• Post-hoc methods on the agentic outputs. Temperature scaling improved VQA-RAD ECE from 0.1430 to 0.1165 and MCE from 0.82 to 0.7634; Dirichlet calibration improved ECE from 0.1437 to 0.0410 and ACE from 0.1437 to 0.0973, achieving the lowest reported ECE/MCE numbers on VQA-RAD in Table 2 (0.041 and 0.113) but at VQA-RAD accuracy of 65.70%, below the proposed method's 68.20%.
• Dirichlet calibration could not be run on ScienceQA. The paper states this is because probabilities of the other options are unavailable; the ScienceQA cells for Agentic + DC are marked as not available.
• Miscalibration can override majority confidence without debate. In the illustrated failure case, three agents answer "cardinalfish" with confidences 0.85, 0.70 and 0.95 (mean 0.83), while a fourth agent answers "black howler" with confidence 0.90; a naive majority-confidence system would adopt the incorrect, poorly supported answer.
• A discrepancy exists in the paper's own reported figures. Different sections report different ECE/ACE numbers for the same settings — for example, debate on VQA-RAD appears as ECE 0.146 in the contributions and in Table 2 but as 0.143 in the results text, and the contributions state debate yields ScienceQA ECE 0.055 and ACE 0.110, while the results text attributes 0.055 ECE and 0.110 ACE to the AlignCal-fine-tuned agentic setting.

Methodology in Plain English

The method has two stages plus a training objective.

Stage 1 — specialized answer generation. An image-question pair is sent to four VLM backbones (Qwen2.5-VL-3B-Instruct, LLaVA-OneVision, Gemma 3 4B, Phi-4-multimodal-Instruct), each paired with a distinct prompting strategy: Chain-of-Thought for multi-hop reasoning, Self-Ask for recursive decomposition, a Search-style prompt for retrieval cues, and GENREAD-style prompting for structured comprehension. Each agent returns an answer plus a confidence obtained from the geometric mean of next-token probabilities. Because agents phrase the same answer differently, a GPT-3.5 judge merges semantically equivalent answers into K unique stances, each with a frequency f_k and a mean confidence c̄_k.

Stage 2 — generalist debate. Generalist agents are instantiated and assigned a stance sampled proportionally to stance frequency, which keeps a soft bias toward majority views while still allowing minority views to be reconsidered. Each agent argues for its stance, rates rationales on logical consistency, factuality, clarity and conciseness, and uses Chain-of-Verification prompting plus a search-augmented agent to check underlying premises. Each agent then receives one supporting and one opposing argument and produces a final answer with a refined confidence. The final stance is the one with the most supporting agents (majority vote), and the reported confidence is the mean refined confidence of that stance's supporters.

The AlignCal loss. The paper argues that binned ECE understates per-example miscalibration and works with UBCE instead, which averages per-instance gaps between correctness and confidence. Because the correctness indicator is non-differentiable, the authors replace the true conditional probability of correctness with the model's own softmax probability of the ground-truth class, giving L_AlignCal = p_y(1 − p_max) + (1 − p_y)p_max. This is added to focal loss with a weight λ. A worked example in the paper illustrates the difference: five patients all at confidence 0.9 with three correct gives ECE = |0.9 − 0.6| = 0.3 but UBCE = (3×0.1 + 2×0.9)/5 = 0.42. Fine-tuning uses LoRA adapters while the base parameters are frozen.

Evaluation. Calibration is measured with ECE, ACE and MCE, with reliability diagrams; task performance is measured with accuracy, F1, precision and recall. Data: ScienceQA (21,208 multimodal multiple-choice questions) and VQA-RAD (3,515 total visual questions, of which only Yes/No questions are used). Hyperparameters mentioned: label smoothing α = 0.1 and focal loss γ = 2.

Why This Matters

Research impact. The paper positions calibration in VQA as an under-examined problem and argues that no prior work leverages multi-agent methods for calibration in VQA. It couples an agent-orchestration idea (debate) with a training-time loss derived from a calibration upper bound, and reports results for combining both, which contrasts with prior work that usually applies either post-hoc scaling or a training-time fix in isolation.

Real-world applications named in the paper:

• Medical diagnosis, including radiology (the VQA-RAD experiment) and a tumor-screening example where a confidently wrong prediction misleads clinicians.
• Autonomous navigation, where a misjudged scene description could trigger hazardous decisions.
• Assistive technologies for the visually impaired, where a high-confidence false answer reduces user trust and poses safety risks.
• Science education question answering (the ScienceQA benchmark), where a system that knows when it is unsure can defer rather than assert.

Industry relevance. Any deployment that lets a VLM act autonomously needs a trustworthy confidence signal for routing, abstention or human escalation. The paper's framework is described as model-agnostic — any set of VLM backbones can be substituted — and the training recipe uses LoRA, which keeps the adaptation cost low relative to full fine-tuning.

Future Directions

• Overcoming the Dirichlet calibration limitation. Dirichlet calibration could not be applied to ScienceQA because probabilities for the non-selected options are unavailable; producing or recovering such probabilities would allow a like-for-like comparison with the proposed loss on that benchmark.
• Tightening worst-case guarantees. The paper notes that worst-case deviation (MCE) is not guaranteed by minimizing an expectation, because no extra uniformity constraint is imposed, even though practical improvements in MCE were observed.
• Scaling the debate. The supplementary material contains ablations on the number of agents and the number of debate rounds; how far these can be scaled before accuracy or calibration degrades, and at what computational cost, remains open.
• Extending beyond the restricted answer settings. Evaluation is limited to multiple-choice VQA and to Yes/No questions in VQA-RAD, leaving open-ended VQA and the broader radiology question types unaddressed.

Target Audience

Researchers and practitioners working on trustworthy multimodal AI, model calibration and uncertainty quantification, and agentic LLM/VLM systems. It is also relevant to engineers deploying VQA or VLM pipelines in regulated or safety-critical domains (medical imaging, robotics, accessibility tools) who need confidence scores they can act on, and to readers interested in how training-time loss design and inference-time multi-agent deliberation can be combined rather than used separately.

Authors’ 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.