UTTBench: Benchmarking Large Multimodal Models for text recognition in underwater thermal turbulent environments

The rapid advancement of Large Multimodal Models (LMMs) has significantly expanded their potential for complex real-world applications. However, their effectiveness in extreme physical conditions, such as underwater thermal turbulence, remains understudied due to the lack of standardized evaluation benchmarks. To address this gap, we introduce the Underwater Thermal Turbulence Benchmark (UTTBench), the first comprehensive benchmark designed to evaluate text recognition in underwater thermal turbulent environments. We conduct a detailed evaluation of four popular LMMs, including LLaVA-Onevision, Qwen2.5-VL, InternVL 2.5, and DeepSeek-VL2, on this benchmark. Our experiments reveal that even advanced LMMs face substantial challenges in accurately recognizing text under thermal turbulence. This study underscores the critical need for further research to enhance the robustness and reliability of LMMs in such challenging environments.