Instinct vs. Reflection: Unifying Token and Verbalized Confidence in Multimodal Large Models

2026-04-19

AI Summary

This work addresses the unreliable confidence estimation of multimodal large language models (MLLMs) in deployment, which stems from a misalignment between implicit token-level support and explicit linguistic expressions of confidence. The study is the first to uncover and systematically analyze the discrepancy between two internal confidence signals—termed “intuitive” and “reflective” pathways—and proposes a monotonic confidence fusion framework. This framework estimates answer correctness through cross-channel consistency modeling and corrects global bias via isotonic mean alignment, thereby achieving well-calibrated selective prediction while preserving the trade-off between risk and coverage. Evaluated across multiple open- and closed-source MLLMs, the method significantly improves confidence reliability, calibration performance, and the ability to anticipate prediction failures.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Natural Language Processing: Language Grounding & Multi-modal NLP
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Multimodal Large Language Models (MLLMs) have demonstrated exceptional capabilities in various perception and reasoning tasks. Despite this success, ensuring their reliability in practical deployment necessitates robust confidence estimation. Prior works have predominantly focused on text-only LLMs, often relying on computationally expensive self-consistency sampling. In this paper, we extend this to multimodal settings and conduct a comprehensive evaluation of MLLMs' response confidence estimation. Our analysis reveals a significant instinct-reflection misalignment: the model's implicit token-level support frequently diverges from its verbal self-assessment confidence. To address this misalignment, we propose a monotone confidence fusion framework to merge dual-channel signals and cross-channel consistency to estimate correctness. Subsequently, an order-preserving mean alignment step is applied to correct global bias, which improves calibration while preserving the risk-coverage trade-off for selective prediction. Experiments on diverse open-source and closed-source MLLMs show that our method consistently yields more reliable confidence estimates and improves both calibration and failure prediction. Code will be available at https://github.com/Yunkaidang/Instinct-vs.-Reflection.

Problem

• Multimodal Large Language Models
• confidence estimation
• instinct-reflection misalignment
• calibration
• selective prediction

Innovation

• confidence estimation
• multimodal large language models
• instinct-reflection alignment
• calibration
• selective prediction