École Nationale Supérieure de
Techniques Avancées

Résumé

I was able to complete my internship at BMW Group in FIZ (Forschungs- und Innovationszentrum) which is the research and innovation center of BMW in Munich. My main work focused on developing a multi-agent system for internal information retrieval tasks and question answering using Large Language Models (LLMs). The objective was to define the architecture of such system and enabling it multi-modal capabilities. From defining how agents should interact together to defining the tools they have access to interact with the real world, this subject is deeply rooted in one of the main challenges that we encounter today with LLM agents: How to make them reason and tackle a problem logically? Indeed, while LLMs have amazing translation capabilities (i.e., natural language to SQL), they lack a fundamental skill which is reasoning and consistently explaining their responses in a logical way. In fact, LLMs are trained to effectively predict sequences of words that have a high probability to appear together, and so in the current state, the reasoning they are capable of is only an illusion of the demonstrations present in their training corpus. This is why they might perform badly on very simple problems that were not present in their training corpus. To put it in a nutshell, LLMs are inherently uncapable of reasoning but excel in giving us the illusion they can. Therefore, my work aimed at finding how we can use different LLM agents and prompts to interact together to get the best reasoning capabilities and tackle efficiently problems. In a second phase, my work was also to make a proof of concept of such a system by developing a user interface with BMW’s frontend framework in Angular.