SEA: Search Engines Amsterdam

This session focuses on Reasoning for Search Systems, featuring two speakers: Alex Egg from Adyen and Mohanna Hoveyda from Radboud University.

Location: Science Park 904, Room C3.161
Date: Friday, March 27
Time: 16:00-17:00
Zoom link: https://uva-live.zoom.us/j/65011610507

Details below:

Speaker #1: Alex Egg from Adyen

Title: Dense Retrieval for Identity: Real-Time Entity Linking in Payments

Abstract: Modern payment platforms must link billions of transactions to the correct shopper identity in real time. Traditional solutions rely on rule-based logic over identifiers such as card hashes, emails, or device fingerprints, but these approaches struggle when identifiers are incomplete, inconsistent, or shared across multiple individuals. In this talk, we describe how we reformulate customer event linking as a dense retrieval problem, where transactions act as queries and shopper identities act as documents in a vector search system. Transactions are encoded using a pretrained neural embedding model fine-tuned with contrastive learning, and incoming transactions are matched to shopper identities via approximate nearest neighbor (ANN) search over billions of shopper embeddings. We will discuss the representation learning challenges that arise in this setting, including learning identity signals from partially observed identifiers and aggregating transaction histories into shopper representations, as well as the system architecture required to support real-time retrieval at scale using hierarchical routing, vector compression, and distributed ANN search.

Bio: Alex Egg is a Research Scientist at Adyen with a background in reinforcement learning. He previously worked on RL and LLM research at Meta, eBay, and UCSD.

Speaker #2: Mohanna Hoveyda from Radboud University

Title: The Reasoning Gap in Search: Overview of Current Paradigms and a Neurosymbolic Solution

Abstract: Information retrieval has long focused on ranking documents by semantic relatedness. Yet many real-world information needs demand more: enforcement of logical constraints, multi-step inference, and synthesis of multiple pieces of evidence.
Meeting these needs turns retrieval into a reasoning problem.
In this talk, we first overview the current landscape of reasoning approaches applied or adaptable to IR from the broader AI research, along with a framework to help compare and relate these approaches for future research.
We then present OrLog, our neuro-symbolic retrieval solution that addresses the reasoning gap for logically constrained queries. OrLog decouples predicate plausibility estimation from logical reasoning: an LLM provides plausibility scores for atomic predicates in a single generation-free forward pass, from which a probabilistic reasoning engine derives the posterior probability of query satisfaction. Evaluated across multiple backbone LLMs and logical constraint types, OrLog significantly improves top-rank precision while reducing token usage by approximately 90% per query-entity pair compared to LLM-as-reasoner baselines.

Bio: I am Mohanna, a PhD student at Radboud University. My research focuses on adaptive orchestration in multi-agent search systems using reinforcement learning, and on developing reasoning methods that require less compute.

Counter: SEA Talks #301 and #302.