Or Levi

Noam Kayzer, Dan Revital, Ori Bar Joseph et al.

We present Hebatron, a Hebrew-specialized open-weight large language model built on the NVIDIA Nemotron-3 sparse Mixture-of-Experts architecture. Training employs a three-phase easy-to-hard curriculum with continuous anti-forgetting anchoring, followed by supervised fine-tuning on 2 million bilingual Hebrew--English samples. The curriculum ordering alone yields a 3-point aggregate benchmark gain over the reversed configuration. Hebatron achieves a Hebrew reasoning average of 73.8\%, outperforming DictaLM-3.0-24B-Thinking (68.9\%) and remaining competitive with Gemma-3-27B-IT on GSM8K-HE and Israeli Trivia, while activating only 3B parameters per forward pass across a 30B-parameter model, delivering approximately 9 times higher inference throughput at native context lengths up to 65,536 tokens. To our knowledge, this is the first language-specific adaptation of the Nemotron-3 architecture for any target language, and the first open-weight Hebrew-specialized MoE model with native long-context support. Model weights are released openly to support further research in Hebrew and Semitic-language NLP.

Uriel Singer, Haggai Roitman, Yotam Eshel et al.

In e-commerce, the watchlist enables users to track items over time and has emerged as a primary feature, playing an important role in users' shopping journey. Watchlist items typically have multiple attributes whose values may change over time (e.g., price, quantity). Since many users accumulate dozens of items on their watchlist, and since shopping intents change over time, recommending the top watchlist items in a given context can be valuable. In this work, we study the watchlist functionality in e-commerce and introduce a novel watchlist recommendation task. Our goal is to prioritize which watchlist items the user should pay attention to next by predicting the next items the user will click. We cast this task as a specialized sequential recommendation task and discuss its characteristics. Our proposed recommendation model, Trans2D, is built on top of the Transformer architecture, where we further suggest a novel extended attention mechanism (Attention2D) that allows to learn complex item-item, attribute-attribute and item-attribute patterns from sequential-data with multiple item attributes. Using a large-scale watchlist dataset from eBay, we evaluate our proposed model, where we demonstrate its superiority compared to multiple state-of-the-art baselines, many of which are adapted for this task.

Yotam Eshel, Or Levi, Haggai Roitman et al.

Recent advances in the e-commerce fashion industry have led to an exploration of novel ways to enhance buyer experience via improved personalization. Predicting a proper size for an item to recommend is an important personalization challenge, and is being studied in this work. Earlier works in this field either focused on modeling explicit buyer fitment feedback or modeling of only a single aspect of the problem (e.g., specific category, brand, etc.). More recent works proposed richer models, either content-based or sequence-based, better accounting for content-based aspects of the problem or better modeling the buyer's online journey. However, both these approaches fail in certain scenarios: either when encountering unseen items (sequence-based models) or when encountering new users (content-based models). To address the aforementioned gaps, we propose PreSizE - a novel deep learning framework which utilizes Transformers for accurate size prediction. PreSizE models the effect of both content-based attributes, such as brand and category, and the buyer's purchase history on her size preferences. Using an extensive set of experiments on a large-scale e-commerce dataset, we demonstrate that PreSizE is capable of achieving superior prediction performance compared to previous state-of-the-art baselines. By encoding item attributes, PreSizE better handles cold-start cases with unseen items, and cases where buyers have little past purchase data. As a proof of concept, we demonstrate that size predictions made by PreSizE can be effectively integrated into an existing production recommender system yielding very effective features and significantly improving recommendations.

Lily Li, Or Levi, Pedram Hosseini et al.

Satire is a form of humorous critique, but it is sometimes misinterpreted by readers as legitimate news, which can lead to harmful consequences. We observe that the images used in satirical news articles often contain absurd or ridiculous content and that image manipulation is used to create fictional scenarios. While previous work have studied text-based methods, in this work we propose a multi-modal approach based on state-of-the-art visiolinguistic model ViLBERT. To this end, we create a new dataset consisting of images and headlines of regular and satirical news for the task of satire detection. We fine-tune ViLBERT on the dataset and train a convolutional neural network that uses an image forensics technique. Evaluation on the dataset shows that our proposed multi-modal approach outperforms image-only, text-only, and simple fusion baselines.