Reproduce “DNA: Unlocking Depth Scaling in CTR Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Caiyuan Li, Wenjie Wu, Yuheng Huang, Xiao Zhou, Silu Zhou, Ke Chen, Zinan Zhao, Xinrui Zeng and 9 more

The authors propose DNA, a long-sequence click-through-rate model that lets request features update behavior representations between cross-attention layers, then shares the updated sequence across candidates. On Taobao search, at about 53 GFLOPs per sample, DNA-10 gains 0.565 AUC points over CrossAttn-4, versus 0.306 for CrossAttn-28, a matched-compute advantage of 0.259 points; this is offline evidence, not online CTR.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.