# Real-Time Hard Negative Sampling via LLM-based Clustering for Large-Scale Two-Tower Retrieval

arXiv: https://arxiv.org/abs/2607.00448 · HTML: https://mlsanity.com/paper/2607.00448/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 17 of its 34 references that are in the map.
- First posted on arXiv 1 Jul 2026.

## Builds on (most cited first, 17 of 17)

- [Session-based Recommendations with Recurrent Neural Networks](https://mlsanity.com/paper/1511.06939/) (2015)
- [Bias and Debias in Recommender System: A Survey and Future Directions](https://mlsanity.com/paper/2010.03240/) (2020)
- [Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations](https://mlsanity.com/paper/2402.17152/) (2024)
- [Recurrent Neural Networks with Top-k Gains for Session-based Recommendations](https://mlsanity.com/paper/1706.03847/) (2017)
- [Embedding-based Retrieval in Facebook Search](https://mlsanity.com/paper/2006.11632/) (2020)
- [Learning Dense Representations for Entity Retrieval](https://mlsanity.com/paper/1909.10506/) (2019)
- [Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System](https://mlsanity.com/paper/2010.15363/) (2020)
- [Managing Popularity Bias in Recommender Systems with Personalized Re-ranking](https://mlsanity.com/paper/1901.07555/) (2019)
- [IRGAN: A Minimax Game for Unifying Generative and Discriminative Information Retrieval Models](https://mlsanity.com/paper/1705.10513/) (2017)
- [Controlling Fairness and Bias in Dynamic Learning-to-Rank](https://mlsanity.com/paper/2005.14713/) (2020)
- [Computationally Efficient Optimization of Plackett-Luce Ranking Models for Relevance and Fairness](https://mlsanity.com/paper/2105.00855/) (2021)
- [Simplify and Robustify Negative Sampling for Implicit Collaborative Filtering](https://mlsanity.com/paper/2009.03376/) (2020)
- [Cascade Ranking for Operational E-commerce Search](https://mlsanity.com/paper/1706.02093/) (2017)
- [Cross-Batch Negative Sampling for Training Two-Tower Recommenders](https://mlsanity.com/paper/2110.15154/) (2021)
- [Revisiting Neural Retrieval on Accelerators](https://mlsanity.com/paper/2306.04039/) (2023)
- [Trinity: Syncretizing Multi-/Long-tail/Long-term Interests All in One](https://mlsanity.com/paper/2402.02842/) (2024)
- [KuaiFormer: Transformer-Based Retrieval at Kuaishou](https://mlsanity.com/paper/2411.10057/) (2024)

## Abstract (arXiv)

The two-tower model is widely used in the retrieval stage of large-scale recommendation systems, where training typically relies on in-batch and/or out-of-batch negative sampling. These methods, however, tend to produce easy negatives that the model learns quickly and that provide little training signal. This paper proposes a self-supervised, cluster-based hard negative sampling technique that draws negatives from the same semantic cluster as the positive item; in our production deployment the clusters are derived from large language model (LLM) based multimodal content representations, so that intra-cluster items are genuinely similar and yield informative negatives. To make this deployable at industrial scale, we realize the technique in a real-time, end-to-end framework that maintains a live in-memory item pool and draws cluster-based negatives from it on the fly via global out-of-batch sampling (GOOBS). The framework integrates directly into production two-tower training and serving and scales to billions of training examples with minimal computational overhead. Experiments on four public datasets and a 14-day online A/B test in a large-scale production system show that the proposed technique outperforms widely used industry methods, while also helping to break recommendation feedback loops and substantially reducing popularity bias.