Can AI Bridge the College Application Gap? Alibaba’s Free Volunteer‑Filling Agent Tested by 400K AI Candidates

In recent months the term “Agent” has moved from demo videos to real‑world deployments, and the high‑school college‑choice process in China provides a concrete test case. This year 12.9 million candidates must each fill roughly 50 volunteer slots (up to 112 in Liaoning), choosing from more than 3,000 universities and 2,000 majors within a narrow time window, yet they receive almost no reliable information.

On June 10 Alibaba launched the country’s first full‑cycle, free‑to‑use high‑school volunteer‑filling Agent built on the Qianwen large‑language model. The Agent offers three core capabilities – a volunteer report, a volunteer calendar, and a volunteer Q&A – and operates online from score lookup through admission follow‑up.

In a walkthrough, a Beijing student with a 660 score, a science‑track, and an interest in artificial intelligence opens the system and immediately sees a personalized volunteer calendar listing score release, filing windows and admission milestones. The user then completes a mandatory profile (province, subject, score, rank) which the system converts into a rank using a one‑point‑per‑segment table, followed by optional fields such as target schools, regional preferences, tuition inclination, major preferences, career goals, graduation plans and an MBTI personality test jointly provided with Beijing Normal University. The more detailed the profile, the more tailored the subsequent plan.

After confirming the profile, the Agent generates a comprehensive report that includes school recommendations, major analyses, and a “risk‑mitigation” strategy. The report also offers personalized tips – for example, whether the student should focus on core AI research or supporting fields like mathematics, statistics, or electronic information. An “Adjust” button lets the user modify inputs, prompting the Agent to instantly update the recommendation.

The article explains why an Agent is needed instead of a simple chatbot. A chatbot answers a single query and then stops, whereas an Agent provides three advantages: persistent memory that remembers past user statements, proactive planning through calendar reminders, and the ability to invoke tools and self‑reflect, resulting in more precise answers.

Data from the Qianwen platform shows that 35 % of users only fill the four basic items (province, subject, score, rank) and only 20 % provide more than five conditions, confirming that passive chatbots cannot meet user needs.

Technically, the Agent relies on a knowledge base covering nearly 3,000 universities and over 2,000 majors, enriched with government planning data and top‑enterprise hiring trends. The underlying model was fine‑tuned on the Qianwen LLM using adversarial reinforcement learning and seven reward functions: three objective rewards for fact‑checking, tool usage and stability; three expert‑experience rewards for planning, reflection and result effectiveness; and one reward for output style. Hundreds of volunteer‑filling experts annotate model outputs, evaluate match quality and risk distribution, and feed corrections back into the training loop, iteratively raising the model to expert level.

To validate reliability, the team built a suite of roughly 400,000 “AI candidates” that span the combinatorial space of volunteer choices. These simulated users cover four dimensions – task type, exam simulation, behavior simulation and edge‑case handling – and repeatedly stress‑test the model before real users encounter it, a strategy described as “AI attacks AI.”

The Agent’s architecture consists of three layers – knowledge base, model, and Agent scheduler – each addressing a critical requirement: sufficient data, professional judgment, and dependable execution. Missing any layer would erode trust in the high‑stakes college‑choice scenario.

In conclusion, the volunteer‑filling stage magnifies information inequality, with private tutoring costing thousands of yuan. An AI‑driven Agent can narrow this gap by delivering expert‑level guidance at scale, offering every one of the 12.9 million candidates a fair chance to understand themselves, explore options and plan their future.