| 9:00 | Introduction |
| 9:15 | Invited talk by Louis Milliken (Perplexity) Diffusion-Pretrained Dense and Contextual Embeddings Dense embeddings are essential for web-scale retrieval, but their high memory and storage requirements limit deployment across billions of documents. This talk introduces pplx-embed, a family of multilingual embedding models built on diffusion-based pretraining with multi-stage contrastive learning and native quantization-aware training (QAT), producing INT8 embeddings that achieve competitive retrieval quality with 4× memory and storage reduction. Bidirectional attention enables mean pooling and late chunking to preserve global context across long documents. pplx-embed familiy includes two model variants: pplx-embed-v1 for standard retrieval, achieving competitive results on MTEB, MIRACL, BERGEN, and ToolRet, and pplx-embed-context-v1 for contextual retrieval, setting new state-of-the-art on ConTEB. Evaluations at production-scale on real-world search scenarios validate the effectiveness of our models for large-scale deployment. Bio: Louis is a Member of Technical Staff at Perplexity AI, spending most of his time on training effective embedding models for information retrieval, from data curation to training methodologies. Prior to Perplexity AI, Louis was a member of the COINSE lab at KAIST, South Korea, researching applications of large language models for software engineering in the field of software engineering. |
| 10:00 | Accelerating Personalization Signal Learning via Synthetic Data Improving Search Suggestions for Alphanumeric Queries |
| 10:30 | Coffee Break |
| 11:00 | Invited talk by Johannes Hoffart (SAP) Unlocking Relational Business Data with In-Context-Learning, Moving Toward Operational Knowledge Expectations towards AI in business applications and processes are higher than ever. Business AI has taken a major step forward: foundation models with in-context learning (ICL) now enable rapid adaptation to new enterprise tasks with far less task-specific training, and RPTs (Relational Pretrained Transformers) extend these gains to relational business data. Yet, to make AI truly dependable in daily operations, the next frontier is to move beyond structure and semantics alone and incorporate operational business knowledge: The rules, procedures, controls, and accountability that govern how data is created and how decisions are executed. This keynote will outline a vision for that evolution and introduce the data and core ideas needed to get there: Training business foundation models that combine relational pretraining with representations of operational business knowledge, towards enabling more reliable in-context learning and stronger generalization across business domains. Bio: Johannes Hoffart is the CTO of the AI Unit at SAP, leading a group of technology experts and scientists driving the research and development of business foundation models and knowledge graphs on SAP's structured data. Before joining SAP in 2021, Johannes has led an AI research group on NLP and Knowledge Graphs at Goldman Sachs and co-founded a spin-off from the Max Planck Institute for Informatics with the goal of enabling businesses to tap into their knowledge hidden in text. |
| 11:45 | Understanding Multi-Structured Documents via LLMs Shivani Upadhyay, Messiah Ataey, Syed Shariyar Murtaza, Yifan Nie and Jimmy Lin Exploring Neural IR in Europeana Suhaib Basir, Mónica Marrero and Julián Urbano Behavioural Effects of Agentic Messaging: A Case Study on a Financial Service Application Olivier Jeunen and Schaun Wheeler |
| 12:30 | Lunch |
| 13:30 | Invited talk by Antoine Chaffin (LightOn) Encoders and Late Interaction: Open Source Search at LightOn In this talk, we'll cover LightOn's open source work for improving information retrieval performance, through encoder and multi-vector models. We will first introduce our effort on creating updated encoders through the ModernBERT project. We will then compare them to decoders (and decoders repurposed as encoders) on various tasks across various scales in the Ettin study. Having highlighted why encoders are crucial for information retrieval, we will discuss the usual single vector search paradigm failure cases and how late interaction models overcome them. We'll then present PyLate, an open source library extending Sentence Transformers to late interaction models and introduce the state-of-the-art results of models trained with it on various aspect: out-of-domain, long-context, reasoning intensive, and code retrieval, outperforming models up to 45 times bigger. All the models presented are shared publicly with open licenses, alongside code and data to reproduce them. Bio: Antoine is an R&D Machine Learning Engineer currently working at LightOn. During his thesis, he explored guiding generative models to create better synthetic data and train multimodal retrieval models to fight misinformation After joining LightOn, he has focused on Information Retrieval, notably by co-leading the ModernBERT project and co-creating PyLate, a library to train and experiment with multi-vector retrieval, which lead to state-of-the-art models such as GTE-ModernColBERT, Reason-ModernColBERT, LateOn-Code and ColBERT-Zero. Antoine also continues to work on multimodal projects, notably by the creation of OCR-free retrieval pipelines and visual document rerankers such as MonoQwen. |
| 14:15 | Evalugator - Rapid, Agile Development and Evaluation of Retrieval Augmented Generation Systems Without Labels Bevan Koopman, Hang Li, Shuai Wang and Guido Zuccon A Systematic Analysis of Chunking Strategies for Reliable Agentic Question Answering Sofia Bennani and Charles Moslonka Iterative Reranking as a Compute-Scaling Method for LLM-based Rankers Tamara Czinczoll, Dong Liu and Filippo Betello |
| 15:00 | Coffee Break |
| 15:30 | Invited talk by Gonzalo Fiz Pontiveros and Roger Zhe Li (Huawei) Generative Recommendation & Multi-Modal Semantic IDs: Industrial Insights from Huawei The rapid advancement of large language models (LLMs) is reshaping how recommender systems are designed and deployed. Benefiting from the powerful generative capabilities of LLMs, large-scale industrial recommender systems are shifting from traditional multi-stage cascaded architectures toward a unified one-step generation paradigm. Additionally,multi-modalsemanticIDs-anaturalby-productofgenerative recommendationmodels-canserveaslow-dimensional,information-rich features for downstream tasks, helping to address real-world challenges such as cold-start problems and popularity bias. By compressing semantic information from user interactions, item content, and contextual signals into ID-like representations, these features enable efficient and low latency use in production systems. In this presentation, we share our perspective on generative recommendation and multi-modal semantic IDs, highlighting best practices and lessons learned from our industrial explorations. |
| 16:15 | Data Augmentation with LLMs for Cold Start Recommendation in E-Commerce Natalija Glisovic, Martin Tegner and Danica Kragic Display Ads Contextual Relevance Modeling with LLM Labels Chao Gan, Fan Yang, Fangping Huang, Weijie Yuan, Nahid Anwar, Musen Wen, Konstantin Shmakov, Hong Yao and Kuang-Chih Lee |
| 16:45 | Closing |
9:00Introduction9:15Invited talk by Louis Milliken (Perplexity)Diffusion-Pretrained Dense and Contextual EmbeddingsDense embeddings are essential for web-scale retrieval, but their high memory and storage requirements limit deployment across billions of documents. This talk introduces pplx-embed, a family of multilingual embedding models built on diffusion-based pretraining with multi-stage contrastive learning and native quantization-aware training (QAT), producing INT8 embeddings that achieve competitive retrieval quality with 4× memory and storage reduction. Bidirectional attention enables mean pooling and late chunking to preserve global context across long documents. pplx-embed familiy includes two model variants: pplx-embed-v1 for standard retrieval, achieving competitive results on MTEB, MIRACL, BERGEN, and ToolRet, and pplx-embed-context-v1 for contextual retrieval, setting new state-of-the-art on ConTEB. Evaluations at production-scale on real-world search scenarios validate the effectiveness of our models for large-scale deployment.Bio:Louis is a Member of Technical Staff at Perplexity AI, spending most of his time on training effective embedding models for information retrieval, from data curation to training methodologies. Prior to Perplexity AI, Louis was a member of the COINSE lab at KAIST, South Korea, researching applications of large language models for software engineering in the field of software engineering.10:00
Accelerating Personalization Signal Learning via Synthetic DataDaraksha Parveen, Doug Kang, Anwitha Paruchuri, Deep Kayal and Pavan Mallapragada
Improving Search Suggestions for Alphanumeric QueriesSamarth Agrawal, Jayanth Yetukuri, Diptesh Kanojia, Qunzhi Zhou and Zhe Wu
10:30Coffee Break11:00Invited talk by Johannes Hoffart ( ... Centrale (Plenary Room) ECIR2026 conference-secretariat@blueboxevents.nlTechnical Issues?
If you're experiencing playback problems, try adjusting the quality or refreshing the page.
Questions for Speakers?
Use the Q&A tab to submit questions that may be addressed in follow-up sessions.
