If we were to sum up what has made humans such a successful species, it would be teamwork. There is growing evidence that when AIs work together, their capabilities can be dramatically improved.
Despite the impressive performance of large-scale language models, companies are still figuring out how to make the most of them. Big tech companies are building AI smarts into a wide range of products, but they haven’t yet found a killer application that will drive widespread adoption.
One of the promising use cases fist The goal is to create AI agents that can perform tasks autonomously. The biggest challenge is that LLMs are still prone to errors, making it difficult to trust complex, multi-step tasks.
But just like humans, two heads seem better than one. A growing body of research on “multi-agent systems” shows that organizing chatbots into teams addresses many of the technology’s weaknesses and enables them to handle tasks that individual AIs cannot reach.
This area saw significant progress from Microsoft researchers last October. A new software library called AutoGen is released Designed to simplify the process of building an LLM team, this package provides all the tools necessary to run multiple instances of LLM-based agents and enable them to communicate with each other via natural language.
Since then, researchers have conducted numerous promising experiments.
In a recent article, mad We highlighted several papers presented at last month's International Conference on Learning Representations (ICLR) workshop. This study shows that performance on math tasks can be improved when agents cooperate. This may be where the LLM struggles, or it may improve reasoning and factual accuracy.
Another example, Mentioned by economistThree LLM-based agents were tasked with defusing bombs in a series of virtual rooms. The AI team performed better than individual agents, with one agent taking on a leadership role and giving orders to the other two, improving team efficiency.
Chi Wang, a Microsoft researcher leading the AutoGen project, told T:he is an economist This approach takes advantage of the fact that most tasks can be split into smaller tasks. LLM teams can process things in parallel rather than sequentially as individual AIs must.
Until now, building a multi-agent team has been a complex process that only AI researchers can really approach. But earlier this month, Microsoft Teams released a new “low-code” interface for building AI teams. Autogen StudioEven non-experts can access it.
The platform allows users to choose from pre-configured AI agents with a variety of characteristics. Or, they can create their own LLM by selecting the LLM that powers the agent, giving the agent “skills” such as the ability to pull information from other applications, and writing short prompts that tell the agent how to behave.
Researchers say that so far, platform users have deployed AI teams for tasks such as trip planning, market research, data extraction and video generation.
However, the approach has limitations. LLMs are expensive to run and can quickly become unsustainable if you let several of them fuss over each other for long. And it's unclear whether AI groups will be more resistant to mistakes, or whether they could cause cascading errors throughout the team.
There is also a lot of work to be done on more mundane tasks, such as how best to structure an AI team and how to distribute responsibilities among its members. There are also questions about how to integrate these AI teams with existing human teams. Nevertheless, pooling AI resources is a promising idea that is quickly gaining momentum.
Image credit: Mohamed Nohassi / Unsplash