An Evening with Linkup Research | On Retrieval

​Linkup Research presents "Field Notes #1 | Retrieval"

3 talks
Q + A
Mixer

Join us at Linkup Paris HQ for an evening on information retrieval, including sparse retrieval, efficient re-ranking, and lessons from shipping production search models.

Thibault Formal, Head of Research at Linkup - “Introducing SPARSEUP: Towards More Competitive Open Sparse Models”

> In this talk, I’ll introduce SPARSEUP, Linkup’s first open-source contribution to the community: a sparse embedding model with under 150M parameters that performs strongly against competing models. SPARSEUP is built on a simple recipe for training SPLADE-like encoders on modern backbones and data, achieving a better efficiency-effectiveness trade-off. In a controlled comparison with the same backbone and data, it comes close to its dense counterpart (DenseOn) while remaining behind late-interaction models (LateOn), which makes it a useful sparse reference point alongside both. Beyond the model itself, I’ll share recent observations that offer a critical view of sparse retrieval: how it behaves on standard benchmarks, and where it runs into limits in large-scale production settings.

Victor Morand, PhD student at Sorbonne Université - “MICE: Minimal Interaction Cross-Encoders for Efficient Re-ranking”

> In this talk, I’ll present our research around MICE, an architecture derived from cross-encoders that aims at keeping their effectiveness, while closing the gap with late-interaction architectures’ efficiency. We’ll first see how cross-encoders can be stripped down into a late-interaction architecture without losing in effectiveness, and then how we managed to turn this into a full retrieval pipeline that uncovers a new region of the Pareto frontier: matching cross-encoder effectiveness while competing with ColBERT-PLAID architecture in efficiency.

Carlos Lassance, Manager of Technical Staff at Cohere - “Shipping Search Models: Lessons from Cohere’s v5 IR Suite”

> In this talk I’ll share my experience developing Cohere’s v5 suite of information retrieval models. We’ll start with training and model merging, and how merging can reveal issues that aren’t visible when looking at individual models. We’ll then look at evaluation and robustness, and what it takes to trust that a model will behave well beyond the datasets it was tested on. From there, we’ll discuss how to choose the “final final” model when different candidates have different strengths and none is best on everything. We’ll finish with data drift: where/when it happens in production is something you can’t really respond to, but can learn from for the next iteration.

Schedule
6:30–7:00 PM - Arrivals
7:00–8:30 PM - Talks and Q&A
8:30–10:00 PM - Mixer with food and drinks

​RSVP here before Tuesday 20th October to secure your spot.