Chinese Teams Win $3 Million at KDD Cup 2026 with DeepSeek and ChatGPT

The KDD Cup 2026 took place at the ICC conference center in Jeju, South Korea, concluding a months‑long algorithm marathon with a total prize pool exceeding $885,000 (≈ 6 million CNY). The academic track champion, team lozyyeah , secured a $200,000 prize, while the industrial track champion, sunshot , earned $100,000.

KDD Cup: A Global Data‑Mining Benchmark

Since its inception in 1997 by ACM SIGKDD, the competition has featured problems set by tech giants such as Google, Microsoft, and Meta. In 2026, Tencent hosted the official TAAC (Tencent Advertising Algorithm Competition) as a KDD Cup main track, providing a real, de‑identified advertising dataset with over 100 feature fields—far larger than any public benchmark.

The dataset’s scale addresses a chronic industry pain point: the scarcity of large, up‑to‑date recommendation system benchmarks, which hampers the study of scaling laws for large language models.

Academic Champion: lozyyeah’s "GPS" for Recommendation Models

Team lozyyeah, led by PhD student Luo, consisted of three members from the same lab. Luo entered the competition seeking a goal‑oriented research problem during the pandemic. The team’s solution, named CRAFT , introduced a novel feature transport mechanism.

Traditional feature interaction methods suffer from information decay across stacked blocks. CRAFT generates an "intent vector" for each layer using flow‑matching ideas, guiding information flow and preserving feature richness until the final prediction. Luo also highlighted the indispensable role of ChatGPT as an AI partner for brainstorming, experiment design, and language polishing of the final paper.

Industrial Champion: sunshot’s QueryFormer

Team sunshot (also called 日之光面 ) won the industrial track with a solution called QueryFormer . Their approach first creates high‑quality queries by deeply interacting user, ad, and non‑sequential features, then retrieves the most relevant information from multi‑domain behavior sequences.

The key insight was that a well‑posed query matters more than longer sequences. QueryFormer implements five optimizations: splitting non‑sequential features by field boundaries, using DCN‑V2 for grouped processing, applying self‑attention then cross‑attention between user and ad tokens, and finally performing cross‑attention‑based retrieval across four business domains. This replaced the original MLP query generator and achieved a test‑set AUC of 0.83254.

Original Breakthrough Awards

The Scaling Law Innovation Award went to the undergraduate team gg , who after 400‑500 experiments discovered that jointly increasing model capacity and conditional information, then adding dimensions, yields scaling improvements.

The Unified Block Innovation Award was granted to team T‑Squared , whose UniBlock architecture unifies disparate modules into a single pipeline, employing a hierarchical memory mechanism with Global Memory for overall summarization and Domain Memory for localized interest extraction.

AI as an Enabler

All award‑winning teams reported extensive use of the DeepSeek web interface for modeling, code development, and paper drafting, while code was written without assistance from code‑generating models like GPT or Claude.

Beyond the Competition: Talent Pipeline

The event serves as a talent discovery platform, feeding participants into Tencent’s long‑term "Qingyun" program for sustained growth. By turning a real industrial problem into a global academic challenge, the competition establishes concrete standards for deployable, industrial‑grade AI solutions and cultivates the next generation of high‑impact AI engineers.