DeepMind’s Talker-Reasoner framework brings System 2 thinking to AI agents.

AI agents must solve a variety of tasks that require different speeds and levels of reasoning and planning abilities. Ideally, the agent should know when to use direct memory and when to use more complex reasoning functions. However, designing an agent system that can handle tasks appropriately according to requirements is still challenging.

at new paperresearcher Google DeepMind We introduce Talker-Reasoner, an agent framework inspired by the “two systems” model of human cognition. This framework allows AI agents to find the right balance between different types of reasoning and provide a more flexible user experience.

System 1 and System 2 thinking in humans and AI

Two-system theory, first proposed by Nobel Prize winner Daniel Kahneman, suggests that human thinking is driven by two different systems. System 1 is fast, intuitive, and automatic. It governs our immediate judgments, such as reacting to sudden events or recognizing familiar patterns. In contrast, System 2 is slow, cautious, and analytical. This enables complex problem solving, planning, and reasoning.

Although often treated as separate systems, these systems continuously interact. System 1 creates impressions, intuitions, and intentions. System 2 evaluates these suggestions and, if accepted, incorporates them into explicit beliefs and intentional choices. These interactions allow us to seamlessly navigate a variety of situations, from everyday routines to difficult problems.

Currently, most AI agents operate in System 1 mode. Excellent for pattern recognition, quick response, and repetitive tasks. However, it is often lacking in scenarios that require the multi-step planning, complex reasoning, and strategic decision-making that are hallmarks of System 2 thinking.

Talker-Reasoner Framework

The Talker-Reasoner framework proposed by DeepMind aims to equip AI agents with both System 1 and System 2 capabilities. We divide the agent into two separate modules: Talker and Reasoner.

Talker is a fast and intuitive component similar to System 1. It handles real-time interaction with the user and the environment. Recognize observations, interpret language, retrieve information from memory, and generate conversational responses. Talker agents typically use the in-context learning (ICL) capabilities of large language models (LLMs) to perform these functions.

The reasoner implements the slow and cautious nature of System 2. Perform complex reasoning and planning. You are equipped to perform specific tasks and interact with tools and external data sources to increase your knowledge and make informed decisions. It also updates the agent’s beliefs as it collects new information. These beliefs inform future decisions and serve as memories that speakers use in conversation.

“Talker agents focus on creating natural, consistent conversations with users and interacting with the environment, while Reasoner agents focus on performing multi-step planning, reasoning, and belief formation based on environmental information provided by the Talker. “I leave it.” Researchers write:

The two modules interact primarily through a shared memory system. Reasoners update their memories with their latest beliefs and inference results, while speakers retrieve this information to guide their interactions. This asynchronous communication allows the Talker to maintain a continuous flow of conversation even while the Reasoner performs more time-consuming calculations in the background.

“This is metaphorical. [the] “A behavioral science dual-system approach, where System 1 is always on and System 2 operates at only a fraction of its capacity,” the researchers wrote. “Similarly, the Talker is always on and interacting with the environment, whereas the Reasoner updates its beliefs, informing the Talker only when the Talker can wait or read it from memory.”

Talker-Reasoner for AI Coaching

The researchers tested the framework in a sleep coaching application. The AI ​​coach interacts with users through natural language to provide personalized guidance and support to improve sleep habits. This application requires a combination of quick, empathetic dialogue and careful, knowledge-based reasoning.

The Talker component of Sleep Coach handles the conversation aspect, providing empathetic responses and guiding users through the various stages of the coaching process. Reasoner maintains the state of your sleep problems, goals, habits, and beliefs about your environment. We use this information to create personalized recommendations and multi-step plans. The same framework can be applied to other applications, such as customer service and personalized training.

DeepMind researchers offer several directions for future research. One of the focuses is optimizing the interaction between Talker and Reasoner. Ideally, the Talker should automatically decide when a query requires the Reasoner’s intervention and when the situation can be handled independently. This minimizes unnecessary calculations and improves overall efficiency.

Another direction is to extend the framework to incorporate multiple reasoners, each specializing in a different type of reasoning or knowledge domain. This allows agents to handle more complex tasks and provide more comprehensive support.