This event is for researchers and engineers working at the intersection of LLMs and recommendation systems. In addition to technical talks from Meta and industry leaders, we're bringing together a community to explore how large language models are reshaping recommendation systems — from research to production.
Join us at Meta HQ in Menlo Park, California for the inaugural LLMs for Recommendation Systems Forum — an afternoon of cutting-edge technical talks, a moderated panel, and a networking happy hour with poster presentations.
The recsys industry is at a familiar inflection point. A decade ago, the shift from linear models to deep learning was met with skepticism — today, the same debate surrounds the move to LLM-native recommendation systems. History shows that companies slow to adapt fall behind. This keynote explores why the convergence is happening now and what it means for the future of recommendations at scale.
Ganesh currently leads prime video recommendations. The team is responsible for all recommendations within the PV app.
He has been working on search, recommendations and growth for ~20 yrs across companies, including LinkedIn, Airbnb, Snap and now Amazon.
Ganesh’s work has been published in conferences including KDD, SIGIR, Recsys etc.
While not involved at work, Ganesh likes to play/compose music and rock climb.
Ashish Rastogi is Director of AI Foundations for Personalization & Recommendations at Netflix, where he leads the team building the core foundation models – including the Netflix member LLM, that power personalization, recommendations, and search for hundreds of millions of members worldwide.
Over nearly a decade at Netflix, Ashish has led teams across content demand modeling, multimodal content understanding, and generative media R&D.
Prior to Netflix, he held machine learning and research roles at Bloomberg, Goldman Sachs, and Google. He holds a Ph.D. in Computer Science from NYU and a B.Tech/M.Tech from IIT Delhi.
Paul's research focuses on adapting and aligning large language models to support creators and audiences in responsible, human-centered ways. His work spans domain adaptation of generative models, preference alignment, and scalable evaluation frameworks that enable models to operate reliably under real-world ambiguity.
Prior to joining Spotify, Paul was a Partner Research Manager in the Augmented Learning and Reasoning group at Microsoft Research.
His research has been published broadly in machine learning and human-centered computing venues and recognized with multiple awards, including a SIGIR Test of Time Award.
Lukasz Heldt works on YouTube’s Recommendation systems, developing SoTA ML applications that power many YouTube recommendation surfaces.
With 11 years of experience, he has proposed and witnessed many technological transformations that fundamentally reshaped how the system operates and optimizes for user satisfaction.
Xinyang Yi is a Principal Research Engineer at Google DeepMind, where he leads research initiatives at the intersection of large language models, personalization, and recommendation algorithms. Over his tenure at Google, his work has focused on advancing the core architectures of retrieval and ranking models, and integrating LLMs into large-scale personalized systems.
His work has led to over 200 launches across multiple key product areas such as YouTube, Play and Ads. Xinyang holds a Ph.D. from The University of Texas at Austin.
Specializing in large-scale AI systems for search, recommendation, and ranking. Over the past two decades, he has built machine learning platforms powering personalized experiences for hundreds of millions of users across LinkedIn and Meta/Facebook. He has authored numerous influential papers on industrial-scale ranking, semantic retrieval, and scalable AI systems.
Fedor is recognized for bridging cutting-edge ML research with production-scale systems, delivering measurable improvements in relevance, engagement, and serving efficiency in real-world applications.
Ying Li is a Staff Research Scientist in the recommendation team at Netflix. At Netflix, her work focuses on large-scale recommendation systems, especially recommendation ranking problems.
Her recent work focuses on transforming the traditional recommendation stack into an LLM‑native recommendation system.Prior to Netflix, she was an Applied Scientist in Amazon, focusing on cold-start classification and large-scale extreme classification using NLP. She obtained her Ph.D. from the University of California, Los Angeles, and B.S. from Peking University.
Luke Simon leads AI research for LLM-RecSys and generative recommender systems at Meta, where his team works on LLM personalization, generative retrieval and ranking, whole-session optimization, and reinforcement learning for ads.
Prior to Meta, Luke was a Distinguished Engineer at LinkedIn focused on LLM-based retrieval and ranking for recommendation and search systems, and Senior Director of ML at Twitter, where he led ads machine learning including targeting, ranking, and auction optimization.
Luke holds a PhD in Computer Science from The University of Texas at Dallas.
Juan's research focuses on pretraining and finetuning foundation models for large-scale personalized systems. His current work studies LLM-based recommenders, integrating catalog knowledge via semantic IDs and modeling long-range user behavior at scale.
He also conducted research at CERN, contributing to the CMS Open Data initiative, and at the Flatiron Institute’s Center for Computational Mathematics, where he worked on online stochastic optimization algorithms. His research has appeared at venues including NeurIPS, ICLR, AISTATS, and ICASSP.
Benyu has over 20 years of experience in architecting large-scale search and recommendation platforms at Facebook, Google, and Microsoft Research. His focus lies at the exact convergence of large language models and personalized systems. A highly cited scholar with over 11,000 citations and 70+ US patents, Benyu’s most recent peer-reviewed research sits at the vanguard of the LLM-RecSys transition.
Benyu brings a uniquely grounded and highly technical perspective to today’s debate on whether LLMs are poised to entirely replace traditional RecSys architectures.
Thea Wang is a Research Engineer at Netflix, where she works on LLM foundation models and generative AI systems for personalization and recommendations.
Her work spans model training, infrastructure and framework development, and enabling the application of foundation models to large-scale recommendation and search systems.
Nikhil earned his MS and PhD from Duke University in Continual Learning of Deep Neural Networks. Nikhil works on superintelligent LLM-based recommenders.
Prior to Meta, Nikhil was a Staff Research Scientist at Google DeepMind, where he led the research efforts for Generative Retrieval and Semantic IDs by pushing the frontiers of personalization and retrieval in Gemini. During his tenure at DeepMind, Nikhil invented key foundational technologies such as GenRetrieval for recommendation systems and Semantic IDs for representing items in a corpus.
Kaushik is working at the intersection of generative AI and recommender systems. His prior work focuses on retrieval scaling, sequential recommendations and graph learning to improve product experiences across Marketplace, Communities, Feed/Home and Monetization. Before joining Meta, he had over 10 years of experience building intelligent products at scale, ranging from job recommendations at LinkedIn and search quality/ranking systems at Google to conversational AI and natural language understanding at Passage AI.
Parish Aggarwal is a Product Manager for the Ranking AI research program at Meta, currently leading the Ads LLM RecSys initiative aimed at building next-generation LLM-powered recommendation systems for ads. Previously, he held product leadership roles at Google Ads, Microsoft AI, and McKinsey & Company.
We share our year-long experience building and operating Rufus, a production shopping agent at Amazon that serves millions of customers. The poster presents our end-to-end agent architecture combining LLM-based reasoning with production-aware retrieval and tool orchestration, a set of system-level optimizations that reduced latency by 45–60%, and a comprehensive evaluation methodology combining human and LLM-as-a-Judge assessment. We demonstrate that LLM-based agents can be reliably deployed at scale for conversational search and recommendation.
In this work, we discuss how Netflix is transitioning the traditional ranking stack to an LLM‑native recommendation system called GenRec. We present our approach of context engineering, post‑training data curation, reward signals, and cost-quality tradeoffs. Despite being trained on 40x less training data, GenRec outperforms the production baseline in both offline analysis and online metrics of a large-scale A/B test. We discuss how LLM-native recommender systems shifts the traditional recommendation systems paradigm in terms of system architecture, feature engineering, modeling strategy, and more.
A widely held hypothesis for why generative recommendation (GR) models outperform conventional item ID-based models is that they generalize better. However, there is few systematic way to verify this hypothesis beyond a superficial comparison of overall performance. To address this gap, we categorize each data instance based on the specific capability required for a correct prediction: either memorization (reusing item transition patterns observed during training) or generalization (composing known patterns to predict unseen item transitions). Extensive experiments show that GR models perform better on instances that require generalization, whereas item ID-based models perform better when memorization is more important. To explain this divergence, we shift the analysis from the item level to the token level and show that what appears to be item-level generalization often reduces to token-level memorization for GR models.
micro1 is a data lab that helps AI labs train foundational models and enterprises develop AI agents. We provide frontier evaluations and reinforcement learning environments to improve LLM capabilities, as well as contextual evaluations to monitor and improve AI agents in enterprise settings. Our data engine includes an AI recruiter agent that sources and vets domain experts, a data platform that enables rapid production of high-quality training data, and a pipeline performance system that ensures both quality and velocity.
We deliver all of this through our managed platform - from expert sourcing and vetting to rubric design, scoring, QA, and structured data outputs - so teams get production-ready training data and evaluation insights without needing to build the operational infrastructure internally.
Ads ranking has long relied on discriminative models that score every candidate ad. We deployed the first end-to-end autoregressive Transformer for ads retrieval at Meta, shifting the paradigm from scoring ads to generating them. This poster covers the ML serving side innovations that made it possible under a strict sub-100ms latency budget: a customized PyTorch-native predictor with multi-stage pipelined inference, a highly optimized decoder with custom SDPA kernels fused into a single CUDA/HIP graph, and in-kernel constrained decoding that guarantees every generated SID maps to valid, eligible ads.
This event is for researchers and engineers working at the intersection of LLMs and recommendation systems. In addition to technical talks from Meta and industry leaders, we're bringing together a community to explore how large language models are reshaping recommendation systems — from research to production.
All registrations are first added to a waitlist and reviewed on a rolling basis. If your registration is confirmed, you'll receive a follow-up email with next steps. Due to limited capacity, attendance is prioritized for researchers and engineers working in the LLM and recommendation systems space. If you're experiencing a technical issue with registration, please contact us here.
For Uber, Lyft, or other ride share services, use 1 Meta Way, Menlo Park, CA 94025 as your drop-off and pick-up location. You can also search "Meta MPK 21.6" in the app — it will appear as a saved stop for Meta.
Shuttles will be available to transport you to the event venue. Return shuttles to the rideshare location will also be available at the end of the event.
Light snacks, along with water and coffee will be provided outside of the presentation room throughout the day and happy hour at the end of the event will have heavy appetizers and drinks.
*Please let us know in the registration form if you have any severe food allergies. Contact us if you have already submitted your registration form.
Ganesh currently leads prime video recommendations. The team is responsible for all recommendations within PV app.
He has been working on search, recommendations and growth for ~20 yrs across companies including LinkedIn, Airbnb, Snap and now Amazon.
Ganesh’s work has been published in conferences including KDD, SIGIR, Recsys etc. While not involved at work Ganesh likes to play / compose music and rock climb.
Ashish Rastogi is Director of AI Foundations for Personalization & Recommendations at Netflix, where he leads the team building the core foundation models – including the Netflix member LLM, that power personalization, recommendations, and search for hundreds of millions of members worldwide.
Over nearly a decade at Netflix, Ashish has led teams across content demand modeling, multimodal content understanding, and generative media R&D.
Prior to Netflix, he held machine learning and research roles at Bloomberg, Goldman Sachs, and Google. He holds a Ph.D. in Computer Science from NYU and a B.Tech/M.Tech from IIT Delhi.
Paul Bennett is Director of Research for LLMs at Spotify, where his research focuses on adapting and aligning large language models to support creators and audiences in responsible, human-centered ways. His work spans domain adaptation of generative models, preference alignment, and scalable evaluation frameworks that enable models to operate reliably under real-world ambiguity.
Paul’s recent research explores how to teach generative systems to reason in structured domains through representations like semantic identifiers, optimize continuously to reflect nuanced human preferences, and generate contextualized narratives that enhance user experience. Across this work he studies how model design and evaluation choices influence alignment with user and creator intent, safety, and controllability.
Prior to joining Spotify, Paul was a Partner Research Manager in the Augmented Learning and Reasoning group at Microsoft Research. His research has been published broadly in machine learning and human-centered computing venues and recognized with multiple awards, including a SIGIR Test of Time Award. His work reflects a commitment to advancing generative AI research that bridges technical depth with human-centered design and safety.
Xinyang Yi is a Principal Research Engineer at Google DeepMind, where he leads research initiatives at the intersection of large language models, personalization, and recommendation algorithms.
Over his tenure at Google, his work has focused on advancing the core architectures of retrieval and ranking models, and integrating LLMs into large-scale personalized systems.
His work has led to over 200 launches across multiple key product areas such as YouTube, Play and Ads. Xinyang holds a Ph.D. from The University of Texas at Austin.
Fedor Borisyuk is a Principal Staff Software Engineer and Technical Lead at LinkedIn specializing in large-scale AI systems for search, recommendation, and ranking. Over the past two decades, he has built machine learning platforms powering personalized experiences for hundreds of millions of users across LinkedIn and Meta/Facebook.
His work spans semantic search, recommender systems, LLM infrastructure, graph neural networks, multimodal learning, computer vision, and efficient large-scale inference. He has authored numerous influential papers on industrial-scale ranking, semantic retrieval, and scalable AI systems.
Fedor is recognized for bridging cutting-edge ML research with production-scale systems, delivering measurable improvements in relevance, engagement, and serving efficiency in real-world applications.
Ying Li is a Staff Research Scientist in the recommendation team at Netflix. At Netflix, her work focuses on large-scale recommendation systems, especially recommendation ranking problems. Her recent work focuses on transforming the traditional recommendation stack into an LLM‑native recommendation system.
Prior to Netflix, she was an Applied Scientist in Amazon, focusing on cold-start classification and large-scale extreme classification using NLP. She obtained her Ph.D. from the University of California, Los Angeles, and B.S. from Peking University.
Benyu Zhang is an AI Research Scientist at Meta, and a veteran AI leader with over 20 years of experience architecting large-scale search and recommendation platforms at Facebook, Google, and Microsoft Research. As the industry shifts toward generative AI, his focus lies at the exact convergence of large language models and personalized systems.
A highly cited scholar with over 11,000 citations and 70+ US patents, Benyu’s most recent peer-reviewed research sits at the vanguard of the LLM-RecSys transition. His 2025 and 2026 publications across ICML, NeurIPS, WWW, and ICDM tackle the field's most pressing bottlenecks—including establishing the first scaling laws for LLMs in recommendation, guiding generative recommenders with human priors, and advancing LLM fine-tuning through Structural Mixture of Residual Experts (S'MoRE).
Having built and scaled traditional, foundational AI systems and now pioneering next-generation LLM approaches, Benyu brings a uniquely grounded and highly technical perspective to today’s debate on whether LLMs are poised to entirely replace traditional RecSys architectures.
Thea Wang is a Research Engineer at Netflix, where she works on LLM foundation models and generative AI systems for personalization and recommendations.
Her work spans model training, infrastructure and framework development, and enabling the application of foundation models to large-scale recommendation and search systems.
Juan Elenter is a Research Scientist at Spotify, where his research focuses on pretraining and finetuning foundation models for large-scale personalized systems. His current work studies LLM-based recommenders, integrating catalog knowledge via semantic IDs and modeling long-range user behavior at scale.
He pursued graduate studies at the University of Pennsylvania, where he focused on optimization and continual learning for large pre-trained models. His theoretical work studied duality-based constrained optimization, showing how dual subgradient methods can yield near-optimal solutions despite non-convexity. He also conducted research at CERN, contributing to the CMS Open Data initiative, and at the Flatiron Institute’s Center for Computational Mathematics, where he worked on online stochastic optimization algorithms. His research has appeared at venues including NeurIPS, ICLR, AISTATS, and ICASSP.
Nikhil earned his MS and PhD from Duke University in Continual Learning of Deep Neural Networks. Nikhil is currently an AI Research Scientist at Meta, where he works on superintelligent LLM-based recommenders.
Prior to Meta, Nikhil was a Staff Research Scientist at Google DeepMind, where he led the research efforts for Generative Retrieval and Semantic IDs by pushing the frontiers of personalization and retrieval in Gemini. During his tenure at DeepMind, Nikhil invented key foundational technologies such as GenRetrieval for recommendation systems and Semantic IDs for representing items in a corpus. His research had significant Google-wide impact, leading to the first GenRetrieval launch at Google and over 50+ subsequent launches—including significant metric gains across YouTube surfaces. His recent work in Generative Retrieval has seen broad industry adoption across several companies including Google, Spotify, Meta, Snap, and Amazon.
Lukasz Heldt works on YouTube’s Recommendation systems, developing SoTA ML applications that power many YouTube recommendation surfaces. With 11 years of experience, he has proposed and witnessed many technological transformations that fundamentally reshaped how the system operates and optimizes for user satisfaction.
Kaushik Rangadurai is a software engineer at Meta working at the intersection of generative AI and recommender systems. His prior work focuses on retrieval scaling, sequential recommendations and graph learning to improve product experiences across Marketplace, Communities, Feed/Home and Monetization.
Before joining Meta, he had over 10 years of experience building intelligent products at scale, ranging from job recommendations at LinkedIn and search quality/ranking systems at Google to conversational AI and natural language understanding at Passage AI.