Imagine teaching an AI to think like a detective, piecing together clues from text and images to solve complex problems. That's the challenge of multi-modal reasoning, a key area of AI research. Large Multi-modal Models (LMMs) are already pretty good at this, but they can struggle with multi-step reasoning. New research introduces ARES, a clever technique that helps LMMs improve their chain-of-thought reasoning, essentially making their thinking process more transparent and logical. Traditionally, AI models learn by being told if their final answer is right or wrong. ARES takes it a step further, providing feedback on each step of the AI’s thought process. It uses advanced AI models like GPT-4 and Claude as "teachers" to give detailed scores on how relevant each sentence of an AI’s reasoning is to the problem at hand. Think of it like a teacher grading each line of a student’s work, not just the final answer. This granular feedback allows the AI to learn which reasoning paths are most fruitful and which lead to dead ends. But that’s not all. ARES also has a second stage where the “teacher” AI corrects specific errors or missing steps in the student AI’s reasoning chain. This correction feedback, combined with supervised fine-tuning, helps the AI learn even faster and avoid getting stuck in bad habits, such as repeating phrases or truncating sentences. The researchers tested ARES on two multi-modal datasets, ScienceQA and A-OKVQA, which involve questions that require understanding both text and images. The results are impressive: ARES consistently generates better reasoning chains than baseline models, as judged by GPT-4, and also improves the accuracy of the final answers. This research opens exciting new avenues for improving multi-modal reasoning in AI. By leveraging the power of advanced AI models as teachers, ARES provides a more nuanced and effective way to train LMMs to think critically and solve complex problems. While there are still challenges, such as dealing with questions that require external knowledge, ARES represents a significant step forward in building AI systems that can reason more effectively about the world around them. Future work will likely focus on enhancing these capabilities further, paving the way for even smarter and more helpful AI assistants.