ProAct: AI that uses idle chat to pre-prepare answers

Shanghai Jiao Tong University and Tencent developed ProAct, an AI agent that uses the quiet time between chat messages to predict likely follow-up questions and prepare answers in advance. The system scans past conversations and stored user data to do background work between user interactions.

ProAct works in three linked stages. The Future-State Prediction stage analyzes prior exchanges, user preferences and missing information to identify probable next questions. The Idle-Time Acquisition stage chooses which predicted needs are worth researching by weighing relevance, timing and expected usefulness. A delivery module then decides whether prepared information should be displayed immediately, saved for later, or kept until needed, forming a closed-loop policy that ties prediction, acquisition and delivery together.

After each interaction the agent updates memory, predicts future needs, allocates idle computation to the most valuable candidates and decides how to handle the resulting preparation, the paper states. The researchers wrote that the system treats idle compute as a constrained resource rather than allowing unconstrained background searches.

The team tested ProAct in 200 simulated sessions across 40 domains, including financial planning, software release management and cybersecurity. In those simulations ProAct reduced the average number of conversation turns by 14.8% and cut follow-up requests by 11.7%. Using a benchmark the researchers call ProActEval, the agent anticipated 703 predictable user needs compared with 32 for an earlier proactive system. The paper reports a 28.1% reduction in hallucinations in the tests.

The study did not involve live users. The researchers identified several limitations: in about 3% of cases ProAct reduced performance by bringing up irrelevant material, and larger Idle-Time Acquisition budgets increased active-token costs while producing diminishing returns. The paper also notes that any production deployment would require privacy protections because the system continuously analyzes conversations and stores user data.

The work appears alongside broader development of persistent, autonomous AI agents that can perform longer or more independent tasks such as coding, scheduling and workflow automation. A UC Riverside doctoral student warned in related research that such agents may pursue goals without fully understanding the consequences and recommended safeguards.

The paper frames ProAct as an operating-point trade-off: the approach can lower user effort and reduce some hallucinations in simulated tests, but it raises computation costs and privacy issues that would need management for real-world use.