About this document
MME-CoT: CoT Benchmark for LMMs by sypxjtu is a document available to read on EtoBox.
The document introduces MME-CoT, a benchmark designed to evaluate the Chain-of-Thought (CoT) reasoning performance of Large Multimodal Models (LMMs) across six domains, including math and science. It highlights key findings, such as the superior CoT quality of models with reflection mechanisms, the detrimental effects of CoT on perception-heavy tasks, and the inefficiency of current models in reasoning processes. The benchmark aims to advance multimodal reasoning by providing a comprehensive evaluation suit
- Author
- sypxjtu
- Language
- EN