In recent years, large language model technology has developed rapidly, but most existing AI agents passively execute instructions and lack initiative. This article introduces a new AI agent jointly developed by Tsinghua University and Face Wall Intelligence. It can predict needs based on user behavior, proactively provide help, and significantly improve user experience. The agent is trained based on a data set called ProactiveBench, which records various user behaviors and is used to train the reward model to determine whether the AI behavior meets human expectations and ultimately realize AI initiative.
In recent years, large language models represented by ChatGPT have set off a new wave in the field of AI. These powerful language models can not only understand human instructions, but also make plans, explore environments, and use tools to solve complex tasks, showing great potential in areas such as robotics, personal assistants, and process automation.
However, most of the existing AI agent systems are passive and require clear human instructions to perform tasks. If you want to schedule a meeting, you have to manually enter the time and location, and even the participants have to be listed one by one. It is simply More troublesome than doing it yourself!
Imagine that you receive an email from a colleague suggesting a meeting, and a passive AI agent waits for you to explicitly instruct it to schedule the meeting. An active AI agent would notice the email and proactively request a meeting. This proactiveness not only significantly reduces the user's cognitive load, but also identifies latent needs that humans have not articulated clearly.
In order to solve the problem of AI assistants being too passive, Tsinghua University and Wall-Facing Intelligence have joined forces to propose a brand new AI agent. It is no longer a machine that "obeys what is told", but can "predict the unknown" before you speak. Before, I took the initiative to help you arrange things clearly!
How does this "magical" AI agent do it? The secret weapon is the ProactiveBench data set! This data set is like an "encyclopedia" that records various human activities, including the information you type in front of the computer. Every letter, every link clicked, and even the content you copied and pasted are clearly recorded!
Using this data set, the researchers trained a reward model, which is like a supercomputer that "simulates the human brain" and can determine whether the AI agent's behavior is in line with human expectations. If the AI agent performs well, it will be rewarded, otherwise it will be deducted points. After repeated training, AI agents can predict your needs based on your behavior just like humans, and proactively provide help when you need it.
For example, when you receive an email from a colleague suggesting a meeting, this "foreseeing" AI agent will automatically identify the content of the email and proactively ask you if you need to schedule a meeting. If you agree, it will automatically help you arrange the time and location, and even send you meeting invitations! Is it much "smarter" than today's AI assistants?
Experimental results show that AI agents trained using the ProactiveBench data set perform very well. For example, the Qwen2-7B-Instruct model has an F1 score of 66.47% in proactively providing help, surpassing all open source and closed source models!
Although this "predictive" AI agent is still in the research stage, it brings new hope for the future progress of human-machine collaboration. I believe that in the near future, we will have an AI assistant that truly "understands you". It can not only "obey you", but can also proactively help you solve various problems, making your life easier and more convenient!
Paper address: https://arxiv.org/pdf/2410.12361
This research result shows the great potential of AI agents to develop in the direction of active services. The application of ProactiveBench data set also provides new ideas for the training of future AI models. I believe that with the continuous advancement of technology, AI assistants will be smarter and more responsive to human needs in the future, and truly become a powerful assistant in our lives.