Proactive Agent: Shifting LLM Agents from Reactive Responses to Active Assistance
o1 Pro.icon
Purpose of the paper.
Conventional LLM agents are mainly "reactive" agents that move after receiving explicit instructions from humans. In this paper, we aim for a "proactive" type agent that monitors user behavior and environmental conditions, anticipates "tasks that the user may need now," and spontaneously suggests tasks. specific approach.
1.Data collection and pipeline construction.
Collect the operation history of actual users in "coding," "document creation," "daily life," etc., and obtain logs (events) of the environment and user behavior from them.
Using a large-scale language model (such as GPT-4), create a virtual "Environment Gym" based on these logs.
On the gym, a large amount of training data is generated by simulating the actions of a user (User Agent), the agent's anticipatory proposals, and the user's acceptance or rejection of them.
2.Building ProactiveBench.
Created a dataset called "ProactiveBench" using the above pipeline (6,790 events in total).
Includes a variety of "event histories" and "anticipatory task suggestions (adopted or rejected)" for training purposes.
A test set (233 events) created from actual user behavior logs is also available.
3.Reward Model.
The human annotator assigns a label to the agent's proposed task as "accept (was necessary)" or "reject (was unnecessary)".
It learns a reward model that mimics this human judgment and automatically judges the "appropriateness (acceptability)" of tasks proposed by the agent.
4.Experiments and evaluations.
We have just learned and fine-tuned based on open source LLMs such as LLaMA and Qwen,
F1 scores for appropriate anticipatory task proposals improved from about 55% to around 66%.
Closed source LLMs such as GPT-4 and Claude were also evaluated, suggesting cases where the fine-tuned models of open source LLMs were able to match or exceed the results.
main results and significance
Denotes a data-driven framework for making [proactive proposals and publishes tools and datasets (ProactiveBench).
By utilizing a compensation model, it is possible to automatically evaluate and improve agents with feedback on whether users really need them.
Experimental results show a steady increase in the accuracy of anticipatory suggestions, but reduction of cases where "suggestions are made in situations where the user does not need them" (False Alarm) and timing adjustment are still issues to be addressed.
future issues.
The simulation environment is limited and needs to be expanded to more diverse scenarios.
The challenge is to understand the context more accurately to reduce the problem of making suggestions even in situations where they are unnecessary for the user.
It is necessary to consider how to integrate with actual user behavior while also taking privacy and ethical aspects into consideration.
--
total
This paper presents a framework for evolving large-scale language models from "reactive" response devices to "proactive" assistants that anticipate and suggest tasks. The main features of the system are a mechanism for creating large amounts of simulation data and the use of that data for fine-tuning and evaluating reward models to enable the agent to read user behavior and suggest task assistance at useful times. While experiments have shown significant performance improvements, the system also presents challenges in actual operation, such as subtlety of timing and frequent false positives.
---