Researchers from Shanghai Jiao Tong University and Tencent have proposed an AI agent system called ProAct, which aims to predict users' next needs before they even speak. Unlike common agents that "wait for questions before answering," this system utilizes the gap between two messages to review historical conversations and saved information, preparing potentially useful content in advance.
Predicting problems using gaps
The research team explained that ProAct's focus is not on faster responses, but rather on turning waiting time into preparation time. The system first predicts the questions the user might ask next based on past conversations, user preferences, and currently missing information.
The system then determines which predictions warrant further processing. Selection criteria include relevance, timeliness, and whether the supplementary information is truly helpful. Once preparation is complete, another mechanism decides whether to display them immediately, save them temporarily, or retrieve them only when the user actually needs them.
The simulation test covers 40 scenarios.
The paper states that ProAct underwent 200 simulation tests across 40 domains, including scenarios such as financial planning, software release management, and cybersecurity. Researchers indicate that this system can identify predictable user needs earlier than previous proactive AI solutions.
- The average number of dialogue rounds decreased by 14.8%.
- Follow-up inquiries decreased by 11.7%.
- Hallucination problems decreased by 28.1%.
In a benchmark test called ProActEval, ProAct identified 703 predictable requirements, while the earlier system identified only 32.
Not yet verified by real users
However, this study remains in a simulated environment. The paper did not include real-user testing, so its effectiveness in actual use remains to be seen. The researchers also acknowledged that in approximately 3% of cases, the system's response quality deteriorated due to the premature presentation of irrelevant information.
Another issue is privacy. Since these agents need to continuously analyze conversation content and store user data, stricter data protection measures are necessary if they are integrated into real products. The paper also mentions that as the backend preprocessing budget increases, the system's token cost rises, but the benefits gradually diminish, meaning that more of this type of "active computation" is not necessarily better.
Competition among AI agents continues to intensify.
This research comes as the tech industry accelerates the development of autonomous AI agents. Related projects are attempting to enable AI to perform tasks such as programming, scheduling, research, and process automation with less human intervention.
At the same time, academics are also warning of the risks associated with such systems. Previous research has warned that some AI agents may perform dangerous tasks without truly understanding the consequences. This means that for proactive agents to be applied in real-world scenarios, in addition to efficiency improvements, safety constraints remain an unavoidable issue.
