Ich habe meine Hochschulaufnahmeprüfungspunkte auf 480 geschätzt, und die KI hat mir 178 Hochschulen ausgewählt. "Mit der bevorstehenden Veröffentlichung der Ergebnisse der jährlichen Hochschulaufnahmeprüfung steht die Frage im Raum: Kann die KI die menschliche Entscheidung treffen?

The annual college entrance examination results are about to be released, and tens of millions of families will face a "complex decision-making process."

During this information-intensive and choice-laden college application process, large language models are becoming a new tool for some parents and examinees.

For example, with Baidu's AI College Application Assistant, by simply entering the examinee's province and score, the AI (Artificial Intelligence) will provide recommendations for three types of college applications: "Reachable," "Safe," and "Guaranteed." A reporter assumed to be a Guangdong examinee who took Physics, Geography, and Politics, and with an estimated college entrance examination score of 480 points, there were a total of 178 colleges to choose from.

AI recommends college application directions based on scores, subjects, and provinces. Image source: Baidu screenshot

However, questions arise: Does AI really understand college applications? And based on what factors can it help users make decisions?

Recently, a reporter from National Business Daily visited Quark, SenseTime, and Baidu to learn about their AI college application tools. Many interviewees said that this year's tools have evolved with the development of large language model technology. They are no longer just search tools that offer a few options but intelligent advisors capable of understanding, reasoning, reflecting, and optimizing plans.

Why has this seemingly "non-commercial" and seasonal field become one of the key projects for technology companies? During the interviews, capability verification, brand building, and information equality emerged as the key concepts.

College application services represent an AI upgrade that requires reasoning in addition to search capabilities

Despite different technical approaches and product forms, the industry has reached several consensuses on the topic of "How to build a college application assistant in the AI era": from information integration to personalized decision-making, and from data add-ons to tool collaboration. In short, the role of large language models is no longer just an "upgraded search" but an intelligent agent with "internalized logic."

"Large language models can now provide comprehensive assistance from information integration to decision support," said the person in charge of Baidu's college entrance examination service project. In college application, large language models can integrate information, provide personalized recommendations, and serve as conversational decision-making assistants.

Baidu has upgraded its "AI Chat (College Application) Intelligent Agent" this year, which can interact with users in natural language and dynamically generate personalized college application recommendations based on the user's score, personality, regional preferences, and other multi-dimensional information. Currently, Baidu Search has officially launched this year's college entrance examination service. By entering "College Entrance Examination" in the Baidu App, users can find the newly upgraded "AI College Application Assistant," college entrance examination big data, and exclusive intelligent agents for more than 2,000 colleges.

On June 12th, Quark released a large language model developed for the college application scenario and launched three core functions: "In-depth College Entrance Examination Search," "College Application Report," and "Intelligent College Selection."

Logical reasoning is the focus of training these models. The college application plans for students aiming for civil service jobs, those interested in the Internet industry, and parents planning for their children to study abroad are completely different.

"A lot of work is needed to teach the model what logic to follow under different conditions," revealed Jia Haifeng, a product manager at Quark. Generating a college application report is equivalent to making tens of thousands of search requests. Combining the examinee's score, subject combination, regional and major preferences, the model will precisely match hundreds of colleges and thousands of majors and generate a well-structured and reasonable personalized recommendation report.

"The logic behind AI's involvement in college application is similar to that of economic or mathematical models. The core is to find the optimal solution based on relevant conditions," explained Jia Anya, the person in charge of the Office Raccoon family business at SenseTime. The newly launched task planning function of SenseTime's Office Raccoon family can guide users to clarify their needs. Then, it will generate execution steps and quickly produce high-quality planning reports. In the college entrance examination scenario, the Office Raccoon allows examinees to input their scores, provinces, preferred majors, or future career directions, and the AI will output a set of "relatively optimal choices" and clearly present the advantages and disadvantages of these choices.

The reasoning process and final recommendations for college applications provided by SenseTime's Office Raccoon tool. Image source: Provided by SenseTime

"The logic of this decision-making assistance essentially uses code to simulate the human brain's decision-making process," said Jia Anya.

She further explained: "In the past six months, the ability of large language models to call external tools has significantly improved, which is the key for AI to handle such complex tasks. Large language models not only have stronger intention understanding and task execution capabilities but also more possibilities to call external tools, such as search modules, score simulators, and employment trend analyzers." This improvement in capabilities enables AI to complete complex decision-making assistance tasks in a phased and orderly manner.

AI is becoming more powerful, but it should not replace human choices

There are two foundations for the effective operation of AI college application assistants: authoritative data and scientific logic.

"Behind large language models are massive amounts of data," said Zhang Fan, the person in charge of Quark Search. Data is the basic requirement for AI to generate "college application reports." This year, Quark's college application large language model has enhanced the construction of the college entrance examination knowledge base by integrating authoritative data from various enrollment colleges and official platforms.

Based on the accuracy of data, understanding the application logic is a unique barrier for college entrance examination large language models. For example, the model needs to understand whether to prioritize colleges or majors.

Jia Haifeng said that a series of processes in Quark's college application report, including reasoning, reflection, execution, and the use of some tools, are only possible due to the improvement of the model's capabilities over the past year. Technically, the model capabilities required for this "reasoning - reflection - execution" structure are higher than before.

Image source: Screenshot of Quark's college application report

Baidu also expressed a similar view. The person in charge of Baidu's college entrance examination project said that as college application is a major event in the lives of all examinees, objective, rigorous, and comprehensive underlying data is the top priority in this field. Limited by the current hallucination problems of large language models and the credibility and comprehensiveness of online search results required for reasoning, large language models' ability to meet the needs of college application highly depends on professional, authoritative, and timely content.

While empowering college application decision-making, multiple platforms have also realized that "large language models should not replace human choices." College application is not a standard question but a complex decision involving multiple dimensions such as values, interests, and career planning.

Zhang Fan said that the platform mainly helps users customize plans based on their personalized needs.

Although AI is very powerful, Jia Anya emphasized the central role of humans in decision-making. "Filling out college applications is a major life decision, and ultimately, humans must be the decision-makers. All tools only play an auxiliary role, which is crucial." She believes that the role of AI is to provide information, analyze pros and cons, and broaden horizons, rather than replacing the examinees' independent choices.

Why do technology companies compete in the "non-commercial" college application field?

The college application scenario has data complexity, decision-making context, and high user participation, making it a natural fit for the implementation and verification of large language models.

"Baidu has been actively involved in college entrance examination services since 2013," said the person in charge of Baidu's college entrance examination project. The driving factors for Baidu's involvement are not only the large number of examinees but also the consideration of promoting information equality and technological equality.

"College application is not just a simple information query. Whether considering the user base, the breadth and depth of the involved information, or the need for personalized information reasoning and integration based on user needs, it is clear that college application is an important scenario for the implementation of large language models," said the person in charge of Baidu's college entrance examination project. Currently, Baidu's "AI College Application Service" has a relatively high user retention rate (users who return the day after their first login) and an average number of conversation rounds compared to most other AI application scenarios, indicating that users have a relatively high level of acceptance and recognition of large language models.

Image source: Screenshot of Baidu's AI Chat for College Applications

Similarly, in Zhang Fan's view, college application is a "very typical information processing scenario" and a scenario where large language models can play their maximum value by improving efficiency and handling complex data relationships. She admitted that although this scenario has a low frequency of use, it is "highly valuable."

Zhang Fan also said that although the user activity of Quark's college application product is not low, Quark always positions it as a public service rather than a commercial product.

Jia Anya believes that complex tasks like college application are a "tough test" for AI capabilities. "If AI can truly help high school students in this decision-making process, they are more likely to continue using our tools in the future when they enter society," she said.

Successfully handling complex problems like college application not only provides practical help to students but also greatly expands the boundaries of tasks that AI can handle. SenseTime's application of its Office Raccoon family products to the college application scenario is not only to demonstrate its AI technology but also to convey the concept of AI empowering the future.

The person in charge of Baidu's college entrance examination project also believes that "Baidu's 'AI College Application Service' may be the first time many high school students deeply interact with and use large language models in real life."

In addition, this experiment of combining AI and education has deeper value. In the college application process, the AI-based conversational interaction will stimulate various new and real user needs. These data and responses will be fed back to the large language model training through a data flywheel.

The seemingly "seasonal" college application assistance product is becoming a brand-building project for AI companies to establish long-term user trust. In the future, AI may participate in more complex and crucial life decision-making moments. And college application service is its first step from the laboratory to the real world.

This article is from the WeChat official account "National Business Daily." Author: Ke Yang, Editors: Cheng Peng, Wen Duo, and Gai Yuanyuan, Proofreader: Zhao Qing. Republished by 36Kr with permission.