Start Date
1-5-2026 12:00 PM
End Date
1-5-2026 1:00 PM
Description
Introduction and Problem Statement
Vision-Language Models (VLMs) are increasingly being integrated into high-stakes environments, such as autonomous robotics used in shared workspaces and as multimodal medical assistant systems. However, current models still have inherent uncertainties, often producing errors or false confidence in responses [1]. Such errors, when passed undetected, can cause serious issues in medical implementations. In shared workspaces, false confidence can lead to safety hazards, or disruption. Du et al. present an evaluation of uncertainty, specifically in the field of human-robot interaction [2]. Such faults are only magnified with the addition of the vision modality, thus highlighting the need for a multi-modal uncertainty quantification, which can identify the source and quantities of uncertainty for each mode.
Proposed Solution
The deficit in the amount of research done in confidence metrics, and uncertainty as a whole, specifically for Vision Language Models, risks undeserving supervised systems, where VLMs are used as assistants in integrated spaces, specifically in collaborative offices with autonomous robots working alongside humans, and as medical assistants. To address the current deficit in uncertainty and confidence quantification for Vision-Language Models, we propose the Modality-Specific Confidence (MSC) framework. It works in a dynamic, three-stage diagnostic pipeline involving stochastic semantic pooling, leading-prompt counterfactual probing, and a mathematical metric to measure linguistic reliance, in the form of a Modality Reliance Score (MRS). An MRS score allows a user to judge the dependence of a model's confidence on visual input as opposed to textual input, thus measuring its resilience to textual prompt injection or sycophancy. By providing a metric for reliability, the MSC framework aims to enhance the trustworthiness of VLMs in collaborative and autonomous workspaces.
Methods and Projected Metrics
We use the stage one confidence score and the stage two confidence score post counterfactual probing to measure the MRS. The counterfactual probing allows us to use a distractor label that tries to change the model’s understanding of the visual input. If that is successful, it suggests a reliance on the language modality of the prompt and lowers true confidence. We use this metric to implement a feedback loop that allows us to measure the effectiveness of the MSC framework in improving accuracy. We also measure overall confidence metrics and check for correlation between them and accuracy.
\[MRS = \frac{CS_{base} - CS_{counter}}{CS_{base}}\]
Impact and Conclusion
Here, we propose both the framework and an evaluation methodology to ensure and measure its effectiveness. To implement the framework, we use a feedback loop that runs a set number of times to conserve resources. We further propose an experiment to calibrate a safety threshold for the MRS score by testing various thresholds and their effect on accuracy. We hope this metric can serve as to improve accuracy and highlight unreliability in VLMs, especially in high-risk scenarios. This research contributes to the broader goal of incorporating VLM assistants in repetitive and diagnostic tasks safely and incorporate risk management.
Modality-Specific Confidence and Uncertainty Metrics for Vision-Language Models
Introduction and Problem Statement
Vision-Language Models (VLMs) are increasingly being integrated into high-stakes environments, such as autonomous robotics used in shared workspaces and as multimodal medical assistant systems. However, current models still have inherent uncertainties, often producing errors or false confidence in responses [1]. Such errors, when passed undetected, can cause serious issues in medical implementations. In shared workspaces, false confidence can lead to safety hazards, or disruption. Du et al. present an evaluation of uncertainty, specifically in the field of human-robot interaction [2]. Such faults are only magnified with the addition of the vision modality, thus highlighting the need for a multi-modal uncertainty quantification, which can identify the source and quantities of uncertainty for each mode.
Proposed Solution
The deficit in the amount of research done in confidence metrics, and uncertainty as a whole, specifically for Vision Language Models, risks undeserving supervised systems, where VLMs are used as assistants in integrated spaces, specifically in collaborative offices with autonomous robots working alongside humans, and as medical assistants. To address the current deficit in uncertainty and confidence quantification for Vision-Language Models, we propose the Modality-Specific Confidence (MSC) framework. It works in a dynamic, three-stage diagnostic pipeline involving stochastic semantic pooling, leading-prompt counterfactual probing, and a mathematical metric to measure linguistic reliance, in the form of a Modality Reliance Score (MRS). An MRS score allows a user to judge the dependence of a model's confidence on visual input as opposed to textual input, thus measuring its resilience to textual prompt injection or sycophancy. By providing a metric for reliability, the MSC framework aims to enhance the trustworthiness of VLMs in collaborative and autonomous workspaces.
Methods and Projected Metrics
We use the stage one confidence score and the stage two confidence score post counterfactual probing to measure the MRS. The counterfactual probing allows us to use a distractor label that tries to change the model’s understanding of the visual input. If that is successful, it suggests a reliance on the language modality of the prompt and lowers true confidence. We use this metric to implement a feedback loop that allows us to measure the effectiveness of the MSC framework in improving accuracy. We also measure overall confidence metrics and check for correlation between them and accuracy.
\[MRS = \frac{CS_{base} - CS_{counter}}{CS_{base}}\]
Impact and Conclusion
Here, we propose both the framework and an evaluation methodology to ensure and measure its effectiveness. To implement the framework, we use a feedback loop that runs a set number of times to conserve resources. We further propose an experiment to calibrate a safety threshold for the MRS score by testing various thresholds and their effect on accuracy. We hope this metric can serve as to improve accuracy and highlight unreliability in VLMs, especially in high-risk scenarios. This research contributes to the broader goal of incorporating VLM assistants in repetitive and diagnostic tasks safely and incorporate risk management.