Benchmarking and Refining Prompting Methods for Visual Logical Reasoning in Multimodal Models

This thesis presents the methodological framework and results of the study focused on the analysis and enhancement of the logical-visual reasoning competencies of MLLMs. The LogicVista benchmark was used to test models on their ability to solve inductive, deductive, and numerical problems where visual input is crucial. The core methodological approach relies on the integration of more advanced prompt engineering techniques, namely Expert Prompting, Chain o Thought and Tree of Thought, implemented with the APET framework. In order to further refine model outputs, specialized RMs, including LLaVA-Critic-7b and DeepSeek-GRM-16B, have been applied within two-stage correction pipelines. The results showed that the optimization of the prompts remains pivotal for larger architectures, while the LLaVA-13B model developed the highest accuracy (75.59%) when combined with the APET approach.