Tencent Advertising Algorithm Challenge 2025: All-Modality Generative Recommendation

A research team from Tencent has introduced a major new benchmark for generative recommender systems (GR) through the Tencent Advertising Algorithm Challenge 2025. The core of the challenge is two novel, large-scale datasets—TencentGR-1M and TencentGR-10M—constructed from de-identified Tencent Ads logs. These datasets are designed to address a critical gap in public research: the lack of realistic, industrial-scale data that includes both collaborative identifiers (like user and item IDs) and rich multi-modal content representations (extracted using state-of-the-art embedding models) for the GR paradigm.

The TencentGR-1M dataset serves as a preliminary track with 1 million user sequences, where each sequence contains up to 100 interacted items labeled with exposure and click signals. The final track, TencentGR-10M, scales this to 10 million users and introduces a more complex, business-critical layer by explicitly distinguishing between click and high-value conversion events. The challenge task focuses on multi-modal sequence generation in an advertising context and employs a weighted evaluation metric that prioritizes predicting conversions. By releasing these datasets and baseline model implementations, Tencent aims to catalyze research into building more powerful and practical generative recommenders that can handle the full complexity of real-world digital advertising.