Click-through rate prediction is a core task in recommender systems and online advertising, where features are categorical and high-dimensional, and their interactions grow combinatorially. A new arXiv paper presents FLMNet, a single-tower model aimed at this setting. The abstract frames the work as a continuation of recent CTR architectures such as FinalNet and QNN.
The method is described in the title as "self-ensemble distillation from masks to seeds." The abstract does not elaborate on the technical details, but the name suggests a training strategy that distills knowledge from masked inputs into seed representations, likely to improve robustness or generalization within a single-tower design.
Because only the abstract is available, the article does not include experimental results or comparisons. The contribution is positioned as a new addition to the growing family of single-tower CTR models that seek to capture feature interactions efficiently.