Extreme Meta-Classification for Large-Scale Zero-Shot Retrieval

A new algorithmic framework called EMMETT aims to improve the accuracy of large-scale retrieval systems when they encounter novel, or zero-shot, items, according to research submitted to arXiv on June 23, 2026 [1]. The framework synthesizes classifiers for new items on-the-fly, addressing a key limitation of existing methods [1]. The research addresses a persistent challenge in information retrieval. Conventional Siamese-style approaches embed queries and items through a single encoder, which allows for the efficient addition of new items but often lacks the capacity for complex tasks [1]. Extreme classification models, which learn a separate classifier for each item in the training set, offer higher accuracy but cannot be trained for novel items due to data and latency constraints [1]. The EMMETT framework, which stands for ExtreMe METa-classificaTion, bridges this gap by synthesizing classifiers for zero-shot items using the readily available classifiers for items seen during training [2]. The paper also introduces a theoretical framework to analyze generalization performance in this setting, which guided the algorithm's design [3]. A specific instance of this framework, called IRENE, is designed for large-scale deployments [1]. In comprehensive experiments, adding IRENE on top of leading encoders improved zero-shot retrieval accuracy by up to 15 percentage points in Recall@10 [1]. The research also moved beyond the lab: an online A/B test within a major search engine's ad retrieval task showed that IRENE improved the ad click-through rate by 4.2% [1]. The authors validated their design choices through extensive ablative experiments and have made the source code for IRENE publicly available [1]. The work builds on the broader machine-learning principle that advances often stem from novel algorithms as well as the availability of high-quality training data [5]. The EMMETT framework leverages millions of observed items with associated user-clicked queries from historical logs to create a massive training set for zero-shot retrieval, ensuring robust generalization on novel items [4]. The paper was shared on arXiv, a platform that, through its arXivLabs project, allows collaborators to develop and share new features while adhering to values of openness and user data privacy [1].