Real-World Industry Problems Take the KDD Stage: Full Highlights of the TAAC 2026 Jeju Island Workshop

On August 12, the TAAC × KDD Cup 2026 workshop was held in ICC‑1, Room 303B on Jeju Island, serving as the closing event of the competition. Participants included teams from UCLA, CUHK, HKU, Microsoft, AppLovin, Google, LinkedIn, Meta and other institutions, representing 52 countries and regions.

The competition, a joint effort between Tencent Advertising Algorithm Competition (TAAC) and the ACM KDD conference, attracted 13,913 contestants forming 5,746 teams and offered a total prize pool of six million RMB. The 2026 challenge, titled “Unified Sequence Modeling and Feature Interaction,” required a single homogeneous recommendation module that simultaneously processes sequential user‑behavior data and heterogeneous non‑sequential features such as user, item, and context attributes. Data were derived from real advertising logs and anonymized. Teams were evaluated on AUC improvement while also considering computational efficiency, inference latency, and deployment practicality, with an emphasis on exploring scaling laws in recommendation systems.

Opening remarks by Irwin King (CUHK) emphasized that the workshop focused on genuine industry‑level technical problems rather than abstract benchmarks, and he encouraged participants to address the concrete issues faced by large enterprises. Rui Li (Meta AI) followed with a discussion on the complexities of scaling recommendation models, highlighting that model growth must be coordinated with system architecture, data pipelines, inference optimization, and training infrastructure.

Fedor Borisyuk (LinkedIn) reviewed LinkedIn’s ranking system evolution from traditional deep models to large language models (LLMs). He detailed techniques such as context compression, model distillation, and system‑level optimizations that enable LLMs to operate efficiently in large‑scale retrieval and ranking scenarios.

Winning solutions illustrated diverse approaches to the unified task:

QueryFormer (industrial track champion “Sunlight”) focused on how the model “asks questions” by employing attention mechanisms to retrieve the most relevant information from behavior sequences for the current task.

CRAFT (academic track champion “lozyyeah”) shifted the perspective from “how features interact” to “how information propagates across layers,” aiming to preserve and organize key information as the model deepens.

CoSFormer (Scaling Law breakthrough award “gg”) argued that when scaling model capacity, new effective information should be introduced rather than merely adding parameters to existing representations.

Other top submissions explored cross‑domain information exchange, token design, capacity allocation, explicit feature interaction, and made trade‑offs among information granularity, compute budget, and internal model connections.

After the main talks, a poster session allowed attendees to dive deeper into model structures, experimental results, and validation methods, extending the discussion from high‑level ideas to concrete design choices.

The workshop concluded with a reaffirmation that “unification” does not imply a single solution; multiple viable strategies will continue to be investigated in future research and industrial practice.