The recent advancements in the field of large multimodal and language models (LMMs and LLMs) have primarily focused on enhancing their mathematical and visual reasoning capabilities. Researchers are developing benchmarks and loss functions to better evaluate and improve these models' ability to handle complex numerical and geometrical tasks. The introduction of benchmarks like TurtleBench, STEM-PoM, and DynaMath underscores the need for more robust evaluation methods to assess the integration of visual understanding and code generation in LMMs, as well as the models' mathematical reasoning robustness. Additionally, novel loss functions, such as those proposed in 'Regress, Don't Guess,' aim to improve numerical accuracy by addressing the limitations of traditional categorical loss functions. These developments highlight a shift towards creating models that can more accurately and reliably process and reason with numerical and visual data, which is crucial for applications in STEM fields. Notably, TurtleBench and DynaMath stand out for their innovative approaches to evaluating visual and mathematical reasoning in LMMs, while 'Regress, Don't Guess' introduces a significant advancement in loss functions for numerical processing in LLMs.