Reproduce “Uncertainty-o: One Model-agnostic Framework for Unveiling Uncertainty in Large Multimodal Models” on OpenResearch

About this paper

Ruiyang Zhang, Hu Zhang, Hao Fei, Zhedong Zheng

Researchers from University of Macau, CSIRO Data61, and National University of Singapore develop Uncertainty-o, a model-agnostic framework that reveals uncertainty in Large Multimodal Models by perturbing multimodal prompts and quantifying semantic consistency of responses, achieving 15.6% higher AUROC on audio comprehension and 9.5% higher AURAC on image comprehension tasks compared to existing baselines through a unified approach that maps diverse multimodal outputs (image, video, audio, point cloud) into text semantic space for entropy-based uncertainty estimation, enabling effective hallucination detection across 18 benchmarks and 10 different LMMs while demonstrating practical utility in self-correction tasks where VideoLLaMA accuracy improved from 62.9% to 67.2% through uncertainty-guided revision and uncertainty-aware Chain-of-Thought reasoning that increased average reasoning steps from 3.1 to 5.4, providing "cookbook" guidelines showing semantic-preserving perturbations and moderate sampling (5 times) optimize uncertainty estimation across both comprehension and generation tasks.

Key insights

• Semantic-preserving perturbations (maintaining core meaning while introducing subtle changes) are more effective than semantic-altering perturbations for revealing true model uncertainty
• Simultaneous perturbation across multiple modalities yields better uncertainty estimation than single-modality approaches
• Progressive pairing of perturbations with similar degrees creates optimal challenge spectrum for LMMs
• Moderate sampling time (5 iterations) balances accuracy without overestimating uncertainty through artificial variance

Summary by alphaXiv.