How Three Undergrads Turned a Last‑Minute Team into a Top‑4 Finish in Tencent’s AI Competition

From a "Buddha‑style" team to rank fourth: formed just two days before the deadline, they relied on calm resilience and score‑driven collaboration to climb the leaderboard. This is the romance of technologists: staying up late tuning parameters, celebrating a 0.1‑point gain, and learning from experts in official communities.

It is hard to imagine that in a Tencent algorithm competition dominated by graduate and post‑doctoral researchers, a team of fresh undergraduates could rise to fourth place.

Their team name, ONLINE_TEST , reflects a desire for continuous improvement, and they never expected to break into the top ten.

When they saw the registration post, only two days remained. Xiao Shun, a classmate, handled computer vision; Xiao Pan, met during graduate school recommendation, had large‑model internship experience; Xiao Ming brought relevant project experience. They quickly formed the team.

The name also meant they hoped each online test would bring progress, and reaching the top ten felt like a victory regardless of the final result.

Initially, they faced a setback: strong offline scores from using graph neural network‑based ID embeddings did not translate to the online leaderboard because many cold‑start users lacked trained ID embeddings.

Through repeated experiments they discovered that introducing overly strong ID features early hindered learning of content features. By adjusting the feature‑introduction order—learning content features first, then adding ID features—they resolved the issue.

They divided the model into orthogonal components, each member focusing on a part, and met weekly to merge progress into a new model.

Disagreements were settled by "letting the scores speak": for example, they tested a user‑click filtering strategy, submitted both versions, and let the online score decide, which proved the filter improved performance.

Having more free time during summer, they could spend hours daily on the competition and relax on weekends, giving them an advantage over graduate competitors.

The community’s collaborative atmosphere helped them; they learned from a top‑ranked participant’s post about a missing logic in negative‑sampling, which, once fixed, nearly doubled their score.

Looking ahead, their goal is to reach the semifinals and, if possible, join Tencent’s advertising recommendation team to apply their technology to real business.

They even joked about using any prize money for a team trip, treating it as a development fund for the next three years.

In the interview’s closing, they shouted their motto: "Promote learning through competition, create great results!"