Reported for NUPIAO |
Edited by NUPIAO | Shuqi Wen
The exams are over, but the real test is just getting started.
Unlike traditional tests with clear-cut answers, filling out college applications is more like an open-ended essay. How do you make a life-altering choice that shapes the next few years, weighing your score, personal interests, city preferences, family background, and career goals across thousands of universities and tens of thousands of majors?
For internet companies, this scenario has become the perfect playground to roll out AI Agents.
Right after this year’s exam season wrapped up, the big players—Alibaba, Tencent, and Baidu—nearly simultaneously upgraded their college prep AI tools. Alibaba’s Tongyi Qianwen rolled out its “College Application Advisor,” positioning it as a full-cycle Agent. Tencent integrated “Yuanbao College Guide” into both its Yuanbao app and QQ Browser, while Baidu supercharged its services with AI-driven recommendation reports. As the LLM race moves into its second half, college admissions have officially become the new battleground for tech giants.
Last year, AI mostly played the role of a basic Q&A bot. This year, the industry consensus is clear: shift from simply answering questions to actively helping users make decisions.
So, when it comes to a decision as massive as college admissions, how does AI actually earn the trust of students and parents alike?
Zheng Sishou, head of product for the Qwen division, shared with NUPIAO that trust isn’t built on just one feature. It’s woven together from accurate data, transparent usage practices, and thoughtful product design.
In his view, AI shouldn’t play God and make choices for you. Instead, it steps back into the role of a trusted advisor or counselor—weighing different options, highlighting pros and cons, and offering directional guidance without just bluntly telling you “Option A is definitely better than Option B.”
“What human experts really bring to the table is proactive communication—helping people figure out what they actually want,” Zheng explained. He noted that right now, not many families hire professional admission counselors, so he hopes to democratize that conversational capability through AI, making it accessible to a much wider audience.
This mindset shift represents the biggest leap for this year’s crop of college AI tools. While last year’s models acted mainly as reference guides, this year, the entire industry is going all-in on autonomous Agents.
Compared to a standard ChatBot, an Agent operates completely differently behind the scenes. According to Jiang Guanjun, chief technology officer at Qwen, traditional LLMs just spit out an answer the moment a prompt hits the server. An Agent, on the other hand, runs a continuous loop: think, plan, execute, reflect, re-plan, and execute again.
Take a practical example: once you plug in your exam score, the Agent won’t just blast out a list of universities. It’ll ask follow-up questions like whether you’re open to moving away from home, if you’re planning to pursue a master’s degree later, or what kind of industries genuinely interest you. Then, it cross-references everything and keeps tweaking the recommendations until it fits.
Jiang believes this mirrors a broader industry trend: large models are evolving from mere conversation partners into actual task-doers. And college admissions? It’s arguably one of the best stress-tests for that exact capability.
The reason is straightforward: nearly every piece of data involved in this process already lives in the digital realm. University stats, enrollment quotas, historical acceptance rates, and major curricula—all of it can be neatly structured. All AI needs to do is sift through that mountain of information and pinpoint the absolute best match for each student.
That said, this doesn’t mean human advisors are about to get replaced entirely.
He points out that humans still hold the emotional edge—they truly empathize with students navigating a stressful transition. Machines, meanwhile, win on breadth of knowledge and raw processing speed.
“A human consultant might need to stop and look up reference materials after making a judgment call,” he added. “AI, powered by LLMs and Agent frameworks, can run dozens of evaluations, rethink strategies, and gather fresh data in just a few minutes—even cross-checking every single line item in a preference list.”
All of this brings us to the real moat: data accumulation has quickly become the foundation of competition among these tech firms.
Speaking for Qwen, they shared that their service leans heavily on Quark’s years of accumulated user data. Last year alone, they used LLMs to generate over ten million personalized reports. This year, they’ve brought in hundreds of domain experts, aiming to hardcode years of hands-on counseling experience directly into the Agent’s logic.
Baidu highlighted that they’ve been riding this wave for over two decades. They’ve spent this year consolidating nationwide university listings, major catalogs, and employment trends, all with the goal of cutting down the decision-making friction caused by information asymmetry.
But let’s be real—in the AI admissions space, raw data is just the entry ticket.
Last year during the exam season, NUPIAO spoke with Liu Jianhua, founder of an independent AI admissions startup called Koubao AI. His take? What really separates a top-tier system from a mediocre one isn’t the underlying LLM—it’s the quality of the backend data and the sharpness of the prediction algorithms.
In his eyes, the holy grail—and biggest technical headache—is cutoff score prediction.
“Theoretically, there’s never just one perfect answer,” he pointed out. The final results don’t even drop until the entire admission cycle wraps up. You’ll see wildly different recommendation lists for the exact same student across different platforms, and that’s usually because they’re running on completely different data architectures and forecasting models.
Jiang doesn’t dodge this elephant in the room either. He openly admitted to NUPIAO that predicting those cutoff lines is undeniably one of the steepest technical cliffs in the whole game.
To crack it, though, you start with one non-negotiable thing: long-term data grinding.
Since China doesn’t have one unified, publicly accessible admissions database, the info is scattered everywhere—from provincial education authorities and university websites to historical enrollment plans and those infamous thick print books students swear by. Building a comprehensive dataset means constantly hunting down, cleaning, and verifying pieces of the puzzle.
“It’s tedious, heavy lifting,” Jiang noted. Just the data-cleaning phase alone creates a massive barrier to entry. Raw inputs are riddled with formatting glitches, missing fields, and flat-out contradictions, demanding serious manpower to cross-check and patch things up.
Liu backed this up, noting that most firms currently have to buy off-the-shelf recruitment archives, run them through OCR scanners, and rely on manual proofreading to build their databases. Some operations even dedicate hundreds of staff members solely to data correction. Plus, since vendors often cross-reference each other’s outputs, a single typo at the source can trigger a domino effect down the chain.
Then comes the second pillar: the prediction algorithm itself.
Jiang explained that raw scores fluctuate wildly from year to year, but admission rankings stay relatively stable. Things like how many spots a university reserves in a specific province or how many seats a major plans to fill are pretty fixed. So, Qwen focuses heavily on mapping score-to-rank relationships, feeding those metrics alongside official quotas and historical books into their algorithms to run predictive simulations.
From Liu’s perspective, legacy systems usually just weight the last three years of admission stats. But in the AI era, you need to feed the model way more dynamic variables—brand-new majors rolling out, universities changing names, sudden shifts in enrollment caps, and the like.
In short, winning this race isn’t about who boasts the biggest parameter count. It’s about who holds the cleanest datasets, the sharpest forecasting engines, and the lowest hallucination rates.
But maybe the question users should actually be asking themselves is: how much decision-making responsibility should we really hand over to AI?
Zheng’s stance is firm: AI is strictly an auxiliary tool, not the final boss making the call. In their design philosophy, the Agent won’t just click the checkbox for you. Instead, it’ll keep asking clarifying questions and interacting with you until you and your parents clearly define your own priorities, then lay out the trade-offs side-by-side.
“Our goal is to help users discover themselves, not to make choices on their behalf,” he emphasized.
He also pointed out that if you wait until twenty days post-exam to start thinking about your interests or career path, you’ve already missed the window. Moving forward, these tools won’t just hang around for that frantic post-Gaokao month. They’ll gradually push upstream, starting as early as freshman or sophomore year of high school.
Baidu’s already experimenting with bridging their exam prep tools directly into career roadmapping and job market trend analysis.
What it boils down to is this: big tech isn’t just fighting over a one-time traffic opportunity. They’re vying for a permanent gateway into long-term life planning.
Just a year ago, everyone was debating whether AI could even handle basic application sorting. Now, with all the tech giants diving in, the playbook has changed. The real question on the table is simpler but harder: whoever understands user intent best, forecasts outcomes most reliably, and builds unshakable trust is the one who’ll step up as the long-term co-pilot for complex life decisions.
(Additional reporting contributed by Wang Qiang)
