Uncertainty Quantification for Multimodal Large Language Models with Incoherence-adjusted Semantic Volume
This addresses the need for reliable uncertainty metrics in MLLMs to enable safe deployment, though it is incremental as it builds on existing uncertainty quantification methods by adapting them to multimodal contexts.
The paper tackles the problem of unreliable outputs from Multimodal Large Language Models (MLLMs) by introducing UMPIRE, a training-free uncertainty quantification framework that computes incoherence-adjusted semantic volume to capture response diversity and incoherence, achieving consistent outperformance over baseline metrics in error detection and uncertainty calibration across various benchmarks and modalities.
Despite their capabilities, Multimodal Large Language Models (MLLMs) may produce plausible but erroneous outputs, hindering reliable deployment. Accurate uncertainty metrics could enable escalation of unreliable queries to human experts or larger models for improved performance. However, existing uncertainty metrics have practical constraints, such as being designed only for specific modalities, reliant on external tools, or computationally expensive. We introduce UMPIRE, a training-free uncertainty quantification framework for MLLMs that works efficiently across various input and output modalities without external tools, relying only on the models' own internal modality features. UMPIRE computes the incoherence-adjusted semantic volume of sampled MLLM responses for a given task instance, effectively capturing both the global semantic diversity of samples and the local incoherence of responses based on internal model confidence. We propose uncertainty desiderata for MLLMs and provide theoretical analysis motivating UMPIRE's design. Extensive experiments show that UMPIRE consistently outperforms baseline metrics in error detection and uncertainty calibration across image, audio, and video-text benchmarks, including adversarial and out-of-distribution settings. We also demonstrate UMPIRE's generalization to non-text output tasks, including image and audio generation.