Confidence Calibration for Multimodal LLMs: An Empirical Study through Medical VQA

Multimodal Large Language Models (MLLMs) are increasingly used in medical diagnostics, yet they often suffer from severe overconfidence—assigning high certainty to incorrect answers. This discrepancy between confidence and accuracy poses significant risks in clinical settings. The authors investigate whether systematic interrogation techniques can better calibrate these models, ensuring that their confidence scores accurately reflect their actual diagnostic reliability.

The study introduces a novel framework called Multi-Strategy Fusion-Based Interrogation (MS-FBI). This system mimics a clinical interrogation process in two phases:

1. Initial Inquiry: The model is prompted to provide an answer and confidence score under a "punishment" constraint, which penalizes high-confidence errors to discourage overconfidence.
2. Deep Inquiry: The model undergoes a dual-track verification process using "Challenge" (logical questioning to expose reasoning flaws) and "Explain" (requiring step-by-step justification) strategies.

After these interactions, an auxiliary expert LLM (Llama3-instruct-8B) reviews the MLLM's initial response, rebuttal, and reasoning to generate a final, calibrated confidence score.

The researchers evaluated this method across three medical VQA datasets (Med-VQA, VQA-RAD, and SLAKE) using four different MLLM backbones. The MS-FBI method consistently outperformed baseline calibration techniques, reducing the average Expected Calibration Error (ECE) by approximately 40%. The study also reveals that medical-domain fine-tuning can sometimes exacerbate overconfidence, and that the optimal combination of interrogation strategies depends on the specific model's pre-training and reasoning capabilities.