# two-tower-model Published articles for two-tower-model. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## Beyond Two Towers: Launching the 3-Tower Engagement Co-Train Model (Part 2) DevFeed: [Beyond Two Towers: Launching the 3-Tower Engagement Co-Train Model (Part 2)]() Original publisher: [Read original article]() Author: Pinterest Engineering Published: 2026-09-17T15:01:05Z Content type: article Language: en Sources: [Pinterest Engineering Blog - Medium]() Topics: [recommendation systems](), [Embeddings](), [Latency](), [Inference](), [AI, ML & Data Engineering]() Tags: [ads](), [architecture](), [cache](), [embeddings](), [engineering](), [latency](), [machine-learning](), [model](), [monetization](), [pinterest](), [ranking](), [recommendation-system](), [scale](), [two-tower-model]() ### AI overview Pinterest engineers describe a three-tower engagement co-train model for lightweight ads ranking. The model jointly predicts click, good-click, and outbound-click probabilities while balancing richer interactions against latency and cost constraints at large candidate volumes. ### Source excerpt Authors: Longyu Zhao (Staff Machine Learning Engineer), Gwendolyn Zhao (Staff Machine Learning Engineer), Peng Yan (Senior Machine Learning Engineer), Yuanlu Bai (Senior Machine Learning Engineer), Yuan Wang (Senior Machine Learning Engineer), Yao Cheng (Staff Machine Learning Engineer), Ang Xu (Principal Machine Learning Engineer), Zhaohong Han (Manager II, Ads Lightweight Ranking) Introduction Previously¹, we launched the next-generation serving stack for standard ads, which we call Nexus. Nexus decoupled candidate generation from scoring and moved us beyond the classic two-tower-only world, enabling richer model architectures while still meeting stringent latency and cost constraints. Building on this system, we set out to design the first ads lightweight ranking model that goes beyond two towers. It jointly predicts three probabilities for each candidate ad: pCTR, the probability of a click; pGCTR30, the probability of a good click that lasts at least 30 seconds; and pOCTR, the probability of an outbound click to the advertiser's destination. To support these objectives efficiently, we partition the query and Pin embeddings into task-specific CTR, gCTR30, and oCTR segments. For the CTR task, the fast two-tower prediction uses the first 64 dimensions of the CTR segment, while the three-tower prediction uses the full CTR segment together with richer cross features. For gCTR30 and oCTR tasks, full embeddings are shared between two-tower and three-tower predictions. This lets each task learn dedicated representations while sharing the overall model. In principle, Nexus places very few hard constraints on the architecture we can serve: cross-attention, sequence modeling, and more expressive interaction modules are all on the table. However, in practice we quickly ran into the fundamental reality of ads lightweight ranking at Pinterest scale: for a typical request, we need to score on the order of hundreds of thousands of candidates (P99 post-targeting candidate count ## Improving Embedding-Based Candidate Generation for Recommender Systems with a Two-Tower Model DevFeed: [Improving Embedding-Based Candidate Generation for Recommender Systems with a Two-Tower Model]() Original publisher: [Read original article]() Author: Ahmad Khan Published: 2026-03-18T21:04:58Z Content type: article Language: en Sources: [Glassdoor Engineering]() Topics: [recommendation systems](), [Machine learning](), [Embeddings](), [Neural Network](), [data](), [ordering]() Tags: [ai](), [data](), [embedding](), [machine-learning](), [model-training](), [neural-network](), [ordering](), [ranking](), [recommender-systems](), [self-supervised-learning](), [two-tower-model]() ### AI overview Glassdoor describes a two-tower neural network for candidate generation in its Community feed recommendation system. The model learns user and post embeddings from interaction data to retrieve potentially relevant posts before filtering, ranking, and ordering. ### Source excerpt Overview At Glassdoor (part of Indeed), millions of professionals come to learn about workplaces, ask career questions, and share experiences through our Community platform. When users open the Community Home Feed, they expect to see posts that are relevant to their interests, industry, and career stage. However, generating that feed is not a simple task. At any moment, thousands of new posts across hundreds of community feeds compete for attention. From discussions about interview experiences to workplace culture questions and industry advice, the platform contains a diverse and constantly evolving stream of content. The challenge is to quickly identify which of these posts are most relevant for each user. To address this problem, we built a multi-stage machine learning recommendation system that efficiently sources, filters, and ranks posts before they appear in the user's feed. In a previous post, Inside Glassdoor's Multi-Stage Recommendation System, we described the architecture of this system and how the different stages work together to deliver personalized content. Snapshot of the Glassdoor Community Home Feed At a high level, the system consists of four stages: Candidate Generation -- retrieves a large set of potentially relevant posts Filtering -- removes posts that are undesirable, duplicated, or already seen Ranking -- predicts which posts a user is most likely to engage with Ordering -- applies heuristics to maintain diversity and freshness in the feed Each stage progressively narrows the set of posts until only the most relevant posts remain. 4 Stage ML RecSys Design In this blog post, we focus on the Candidate Generation stage, which is responsible for retrieving an initial pool of potentially relevant posts from a very large pool of posts across many different feeds. For Candidate Generation, we developed a Two-Tower Neural Network Model that learns embeddings for users and posts based on real interaction data. In the rest of this blog post, we will walk ## Dense Retrieval DevFeed: [Dense Retrieval]() Original publisher: [Read original article]() Author: Laurynas Jasiukėnas Published: 2025-11-18T00:00:00Z Content type: article Language: en Sources: [Vinted]() Topics: [Embeddings](), [Machine learning](), [Fine-tuning](), [rust-ai]() Tags: [business](), [e-commerce](), [embedding](), [experimentation](), [fine-tuning](), [images](), [metrics](), [model](), [retrieval](), [search](), [two-tower-model](), [vector]() ### AI overview Vinted Engineering describes integrating dense embedding-based retrieval into its multilingual e-commerce search system. The approach uses a two-tower model with query and item embeddings, and the article covers experimentation, model improvements, and engineering work that led to full enablement after approximately 50 A/B tests. ### Source excerpt TL;DR: integrating embedding-based retrieval into the e-commerce search application is a significant undertaking.