Just how sure are you? Improving Verbalized Uncertainty Calibration in Medical VQA

Technical Summary: Improving Verbalized Uncertainty Calibration in Medical VQA

Problem Statement
Multimodal Large Language Models (MLLMs) applied to Medical Visual Question Answering (VQA) frequently exhibit overconfidence, generating high-confidence outputs regardless of their actual correctness. This lack of calibrated self-assessment limits their reliability in clinical settings, where clinicians must manually verify reasoning rather than relying on the model's trustworthiness. Existing methods for verbalized confidence calibration are primarily designed for text-only LLMs and fail to account for the multimodal nature of medical image understanding. Furthermore, current approaches often train on single input conditions, failing to teach the model how its confidence should shift when visual evidence or textual integrity is disrupted.

Methodology
The authors propose a training-based framework that fine-tunes MLLMs to produce better-calibrated verbalized confidence scores. The approach relies on three core components:

1. 2 × 2 Factorial Perturbation Design:
   To teach the model to ground confidence in evidence utilization, the authors introduce a perturbation scheme that crosses two axes:

   • Image Availability: Original image vs. a blacked-out image (removing visual evidence without distorting features via noise).
   • Text Integrity: Original options vs. perturbed options (shuffling order and replacing distractors with options from similar questions to destroy memorized patterns).
     This yields four conditions per question. The model generates multiple responses ($G=10$) for each condition to compute fractional accuracy, which serves as the ground truth target for calibration.

2. Composite Calibration Loss Function:
   The framework optimizes a composite loss consisting of four terms:

   • Brier Loss: Drives the predicted confidence distribution toward the empirical fractional accuracy derived from the perturbation conditions.
   • Anchor Regularizer: Prevents the model from collapsing into extreme confidence values (0 or 1) by penalizing deviations from a midpoint ($c_{mid} = 0.5$) for uncertain samples. This acts as a conservative prior, requiring strong evidence to shift confidence away from maximum uncertainty.
   • Alignment Loss: A contrastive term ensuring that changes in confidence across perturbation conditions track the corresponding changes in empirical accuracy. This forces the model to adjust its confidence based on the informational contribution of specific evidence sources (visual vs. textual).
   • Top-k KL Divergence Regularizer: Applied to the answer token positions (excluding the confidence token) against a frozen copy of the base model. This preserves the model's original answering ability and instruction-following capabilities during the fine-tuning process, preventing accuracy collapse.

3. Training Objective:
   The total loss is a weighted sum of these components. The model is fine-tuned using LoRA adapters, generating a reasoning trace, an answer, and a verbalized confidence token (on a discrete integer scale normalized to [0, 1]) in an answer-first, confidence-second order.

Key Contributions

• First Training-Based Calibration for Medical VQA: Unlike prior work restricted to text-only models or prompt-based engineering, this is the first method to apply training-based calibration specifically to multimodal medical models.
• Evidence-Aware Perturbation: The novel 2 × 2 factorial design explicitly probes and teaches the model to distinguish between reliance on visual evidence versus language priors.
• Composite Loss Architecture: The integration of Brier calibration, an anchor regularizer, contrastive alignment, and a KL regularizer creates a robust training signal that simultaneously improves calibration and discrimination without sacrificing predictive accuracy.

Results
The method was evaluated on three Medical VQA benchmarks (OmniMedVQA, PMC-VQA, MedXpertQA) covering high, mid, and low accuracy regimes, using two architectures (MedGemma-4B-IT and Qwen2-VL-7B-Instruct).

• Calibration and Discrimination: The proposed method reduced calibration error (ECE) by 60% or more and improved discrimination (AUROC) by 26% or more compared to baselines, including prompting-based (Top-K Sampling, SteerConf) and training-based (ConfTuner) approaches.
• Robustness: The method maintained consistent performance across different difficulty regimes, whereas baseline ECE degraded significantly as task difficulty increased.
• Accuracy Preservation: The top-k KL regularizer successfully preserved predictive accuracy, with Out-of-Distribution (OOD) accuracy changes remaining within ±0.8 percentage points.
• Ablation Studies: Removing the alignment term caused a sharp drop in AUROC (discrimination), while removing the KL regularizer led to significant accuracy collapse and a loss of instruction-following capabilities (e.g., failure to output confidence in the correct format).

Significance and Claims
The paper claims that MLLMs can learn meaningful verbalized confidence through targeted fine-tuning that explicitly models the relationship between evidence sources and prediction correctness. The authors assert that their method achieves a simultaneous improvement in both calibration (stated confidence matching empirical correctness) and discrimination (correct predictions receiving higher confidence than incorrect ones), a trade-off often observed in prior work.

The authors maintain a modest stance regarding limitations:

• Discrimination Limits: On the most challenging benchmark (MedXpertQA, ~20% base accuracy), no method achieved meaningful discrimination, suggesting verbalized confidence cannot yet distinguish correct from incorrect predictions in tasks requiring complex multi-step clinical reasoning.
• Scope: The work is limited to multiple-choice VQA and small-scale models (4B and 7B parameters). The authors note that extending these gains to open-ended VQA and larger models remains an open question.
• Clinical Deployment: While the method improves reliability, the authors emphasize that calibrated confidence is intended to support, not replace, human judgment, and further validation in real clinical conditions is required before deployment.