Uncertainty Quantification for Multimodal Large Language Models with Incoherence-adjusted Semantic Volume

Researchers introduce UMPIRE, a new training-free framework for quantifying uncertainty in Multimodal Large Language Models (MLLMs) across various input and output modalities. The system measures incoherence-adjusted semantic volume of model responses to better detect errors and improve reliability without requiring external tools or additional computational overhead.

• →UMPIRE provides uncertainty quantification for MLLMs across image, audio, and video-text tasks without requiring external tools or additional training.
• →The framework outperforms baseline metrics in error detection and uncertainty calibration across multiple benchmarks including adversarial scenarios.
• →UMPIRE works by computing semantic volume of sampled responses while adjusting for local incoherence based on internal model confidence.
• →The system generalizes to non-text output tasks including image and audio generation applications.
• →This advancement could enable better deployment of MLLMs by identifying when to escalate unreliable queries to human experts or larger models.