Meet Flows: A Revolutionary AI Framework for Modeling Complex AI-Human Interactions

Recent advancements in artificial intelligence have created multiple opportunities for structured reasoning as they can remarkably adapt to information within their context. This collaboration between multiple AI systems and humans is crucial. Strategic content crafting can lead LLMs to perform complex reasoning to enhance their capabilities. We require a principal and organized way of designing and studying such models. EPFL and PSL University researchers propose a “control flows” framework to model complex interactions.

These control flows are tools designed to solve increasingly complex tasks. In simple words, these are self-contained building blocks of computation. These flows can be recursively composed into arbitrarily nested interactions with substantially reduced complexity. Flows represent any collaboration that includes any AI-AI and human-AI interactions. Flows introduce a higher-level abstraction that isolates the state of individual Flows and specifies message-based communication as the only way to interact. Examples of such control flows are ReAct, AutoGPT, and BabyAGI. 

To show the potential of the Flows, researchers selected the task of competitive coding, which involves users trying to solve problems defined by a specification. They designed specific building blocks (flows), which include planning flows, which allowed the AI agents to strategize their approach;  reflective flows, which allowed AI agents to analyze and improve their previous answers; collaborative flows, where one AI agent seeks feedback from another; code testing flows, which involved executing the code and optimizing it based on the results.

They combined these building blocks to create multiple coding flows and evaluated problems taken from CodeForces and LeetCode. Even for advanced models like GPT-4, performing this task is challenging. They found that the GPT-4 solve rate drops to 72%. Whereas their strategy of complex interactions improved the performance, AI-AI interaction’s post-cutoff solve rate by 20%, and human-AI interaction by 54%. 

Researchers claim this framework enables an intuitive and simple design of arbitrary complex interactions. To make this method accessible to all, researchers open-source the ‘aiFlows’ library with a repository of Flows named Flow Verse that can be easily used, extended, and composed into more complex Flows; tools; a detailed logging infrastructure to enable transparent debugging and analysis; a visualization toolkit to examine the Flows’ execution. They also provided detailed documentation and tutorial files to familiarize one. 

Though carefully designing the complex interactions improves generalization, it comes with additional computation and latency costs. Their framework will serve as a solid basis for supporting practical and theoretical innovations in AI, paving a step closer to artificial general intelligence. They say their future work involves building an AI system that can efficiently improve our problem-solving abilities. 

Check out the Paper and Github. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 28k+ ML SubReddit, 40k+ Facebook Community, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.

Arshad is an intern at MarktechPost. He is currently pursuing his Int. MSc Physics from the Indian Institute of Technology Kharagpur. Understanding things to the fundamental level leads to new discoveries which lead to advancement in technology. He is passionate about understanding the nature fundamentally with the help of tools like mathematical models, ML models and AI.

[Announcing Gretel Navigator] Create, edit, and augment tabular data with the first compound AI system trusted by EY, Databricks, Google, and Microsoft