NEWS · RESEARCH · #351
QueryFormer — KDD Cup 2026 Tencent UniRec winner (arXiv:2609.16548v1)
QueryFormer proposes a stackable unified field–sequence block that generates queries via cross-attention and packs sequence queries with shared-parameter attention to bridge non-sequential multi-field features and user behavior sequences. The model won 1st place in the KDD Cup 2026 Tencent UniRec Industrial Track (official test AUC 0.83254; a post-competition scale reached 0.832713), shows H-scaling gains (validation AUC 0.84540→0.84615), outperforms HyFormer at comparable budgets, and keeps H=8 inference latency to 1.89× that of H=1.
KEY POINTS
- QueryFormer proposes a stackable unified field–sequence block that generates queries via cross-attention and packs sequence queries with shared-parameter attention to bridge non-sequential multi-field features and user behavior sequences.
- The model won 1st place in the KDD Cup 2026 Tencent UniRec Industrial Track (official test AUC 0.83254; a post-competition scale reached 0.832713), shows H-scaling gains (validation AUC 0.84540→0.84615), outperforms HyFormer at comparable budgets, and keeps H=8 inference latency to 1.89× that of H=1.
- The paper introduces an attention-based query generation bridge that improves pCVR prediction accuracy and latency trade-offs, validated by a KDD Cup win and controlled scaling/ablation studies.
WHY IT MATTERS
The paper introduces an attention-based query generation bridge that improves pCVR prediction accuracy and latency trade-offs, validated by a KDD Cup win and controlled scaling/ablation studies.