This article introduces a new quantitative investment framework, aiming to mine and optimize the Alpha factor in stock investment strategies through the collaboration of Large Language Models (LLMs) and multi-agent. This method screens and generates a diversified Alpha set from multi-modal financial data (financial reports, prices, news, charts), and then dynamically evaluates and selects the best based on real-time market scenarios through a multi-agent mechanism, thereby building a more adaptable and more stable investment portfolio. --- 1. Introduction 1.1 Background and motivation As deep learning has made significant progress in financial transaction prediction, many researchers and practical institutions are working to develop more stable and reliable quantitative investment models. However, current models often perform erratically in different market environments, and their predictive capabilities are susceptible to noisy data or market fluctuations. In addition, in real trading scenarios, data comes in various forms (text, charts, financial indicators, etc.). How to better integrate and refine this information has become a new challenge for quantitative investment. The core ideas of the new quantitative investment framework we proposed: (1) Large Language Model (LLM) provides powerful information extraction and reasoning capabilities, and can automatically generate and screen out diverse Alpha factors; (2) Multi-agent system Dynamically…