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

🕒 Published (v1): 2025-11-14 11:08 UTC · Source: Arxiv · Venue: AAAI 2026 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

AlignVQA is a two-stage multi-agent debate framework that improves confidence calibration in VQA by having diverse specialized VLMs generate candidate answers, then having generalist agents critique and revise them via structured debate. A novel differentiable loss (AlignCal) fine-tunes the specialized agents to minimize an upper bound on calibration error, yielding substantially lower ECE when combined with the debate framework.

Problem

State-of-the-art VLMs used for VQA are systematically overconfident—their reported confidence scores do not reflect actual correctness probabilities (high ECE). This miscalibration is dangerous in high-stakes domains (medical diagnosis, autonomous navigation) where downstream decisions depend on reliability of confidence. Prior calibration methods either require retraining, apply post-hoc fixes without iterative refinement, or ignore calibration entirely in multi-agent VQA architectures.

Method

Stage 1 — Diverse Expert Generation: Four VLM backbones (Qwen2.5-VL-3B, LLaVA-OneVision, Gemma 3 4B, Phi-4-multimodal) each run with a distinct prompting strategy (Chain-of-Thought, Self-Ask, Search-Augmented, GENREAD). Outputs are semantically clustered into K unique stances using GPT-3.5 as a judge, yielding per-stance frequency and mean confidence.

Stage 2 — Generalist Debate: Generalist agents are instantiated, each probabilistically assigned a stance proportional to Stage 1 frequencies. Agents argue for their stance, receive Chain-of-Verification-style for/against feedback from all other agents (factuality-checked via search), then revise their answer and confidence. Final prediction is majority vote; final confidence is mean confidence of supporting agents.

AlignCal Loss: A differentiable surrogate loss that minimizes the Upper Bound Calibration Error (UBCE). Given softmax output \(p_y\) (true class probability) and \(p_{\max}\) (top confidence): $\(\mathcal{L}_{\text{AlignCal}} = p_y(1 - p_{\max}) + (1 - p_y)p_{\max}\)$ Combined with focal loss (\(\mathcal{L}_{tot} = \mathcal{L}_{FL} + \lambda \mathcal{L}_{AlignCal}\)), this fine-tunes the specialized agents. The gradient drives \(p_y \to 1\) when correct and \(p_{\max} \to 0\) when wrong.

Key Contributions

• AlignVQA framework: two-stage agentic debate (specialist generation + generalist deliberation) for post-hoc calibration in MCQ VQA, model-agnostic over VLM backbones.
• AlignCal loss: closed-form differentiable surrogate to UBCE; hyperparameter-light (only \(\lambda\)); no kernel choice or binning required unlike MMCE.
• Demonstrates that combining calibration-aware fine-tuning with the debate framework compounds improvements beyond either technique alone.

Results

• VQARad (Gemma 3 4B baseline ECE = 0.375):
• Agentic framework alone: ECE → 0.146, ACE → 0.133
• AlignCal + FL fine-tuning alone: ECE → 0.137, ACE → 0.115
• Agentic + AlignCal + FL (proposed): ECE → 0.098, ACE → 0.095, MCE → 0.267
• ScienceQA (Gemma 3 4B baseline ECE = 0.396):
• Agentic framework alone: ECE → 0.270, ACE → 0.265
• AlignCal + FL fine-tuning alone: ECE → 0.058, ACE → 0.120
• Agentic + AlignCal + FL (proposed): ECE → 0.055, ACE → 0.110, MCE → 0.331
• Outperforms post-hoc alternatives (temperature scaling, Dirichlet calibration) and train-time alternatives (focal loss, label smoothing) in ECE/ACE on at least one dataset.
• Accuracy is competitive but not always best (e.g., 68.2% vs 75.2% for LS on VQARad).

Limitations

• Accuracy can degrade relative to single-model baselines or simpler fine-tuning approaches (e.g., LS achieves 75.2% on VQARad vs 68.2% for the proposed method).
• MCE improvements are not theoretically guaranteed by AlignCal (minimizes an expectation, not the worst-case bin gap); only empirically observed.
• Evaluated only on two datasets (ScienceQA, VQARad Yes/No subset), limiting generalization claims.
• Dirichlet calibration cannot be applied to ScienceQA due to unavailability of other-option probabilities, making post-hoc comparisons incomplete.
• Computational cost of the two-stage multi-agent debate (4 specialized + multiple generalist agents + GPT-3.5 judge) is not analyzed.

Relevance to Agentic AI / LLM Agents

This paper directly demonstrates a practical use of multi-agent debate as a mechanism for improving reliability properties (calibration) rather than just accuracy, which is relevant to agent systems that must act under uncertainty. The two-stage design—diverse specialists followed by deliberating generalists—is a concrete instantiation of role specialization and iterative refinement patterns common in agentic architectures. The finding that calibration-aware fine-tuning of individual agents compounds with debate-level aggregation suggests that the reliability of an agent ensemble depends on both the quality of individual agents and the interaction protocol. For researchers building agentic pipelines in high-stakes domains, this provides an empirically validated path to trustworthy confidence signals without relying solely on post-hoc fixes.