Abstract

Retrieval-Augmented Generation (RAG) significantly improves LLM accuracy by grounding responses in external documents. However, this accuracy often comes at the cost of speed, as longer contexts increase processing latency. 

This talk will share how to apply novel compression techniques to achieve faster RAG—dramatically reducing context length and latency—while maintaining response quality.

The talk will be based on the recent publications:

- Provence:, ICLR 25 https://huggingface.co/blog/nadiinchi/provence

- PISCO, ACL’ 25; https://aclanthology.org/2025.findings-acl.800/

- OSCAR: https://arxiv.org/abs/2504.07109v1 

Short Bio

Stéphane Clinchant is a Principal Scientist and Team Lead currently at Naver Labs Europe. With over 10 years of experience in Machine Learning, Information Retrieval, and NLP , he and his team now focus on retrieval and memory augmented agents models for LLMs. 

Abstract

This talk will explore how optimization powers real-world industry applications. 

We’ll begin with a gentle introduction to different types of optimization problems, then dive into resource allocation through Mixed Integer Linear Programming (MILP). 

Drawing from his experience at Kelkoo, Georgios will present a concrete application in online marketing: using optimization to allocate budgets and drive better business outcomes.

Whether you’re new to optimization or seeking practical case studies, this talk will show why optimization tools deserve a place in every data scientist’s toolkit.

Short Bio

Georgios Balikas leads the Data Science team at Kelkoo, where optimization and machine learning drive thousands of business decisions daily. Previously a Lead Data Scientist at Salesforce, he is also co-organizer of the Grenoble Data Science meetup

Presentation

Georgios kindly shared his presentation in case you missed his talk: Data-Driven Decision-Making with Optimization Lessons from Industry 

Abstract

Agents today are implemented as complex graphs with intricate control systems, requiring engineers to navigate labyrinths of states, transitions, and exception handlers. PREDIBAG (Predicate-Based Agent) challenges this status quo by resurrecting logic programming principles from Prolog's golden era. Where traditional architectures demand explicit path definitions for every scenario, PREDIBAG agents leverage declarative knowledge and inference engines to discover optimal solutions autonomously. 

Engineers trapped in the quagmire of maintenance-heavy graph models will find liberation in PREDIBAG's clean, composable predicates that scale with problem complexity rather than against it. The architecture enables agents to share knowledge seamlessly, reason about their own actions, and integrate Python execution within the logical framework—eliminating the artificial boundary between reasoning and action. 

PREDIBAG doesn't just offer incremental improvements; it fundamentally transforms how we conceptualize agent behavior. In a field desperately seeking solutions to the brittleness of conventional approaches, 

PREDIBAG demonstrates that logic programming isn't just a historical curiosity—it's the overlooked key to unlocking the next generation of robust, adaptive, and truly intelligent agent systems. 

Short Bio

Claude Roux did his PhD at the Université de Montréal and I worked for different companies in AI, starting with INGENIA in Marseille from 1995-1997, then the Xerox Research Centre Europe lab Meylan, which has been acquired by Naver in 2017. 

Claude's main interests are in formal grammar, programming languages and Large Language Models, with a focus on agentification and reasoning. 

Claude has implemented many different softwares over the years such as XIP (Xerox Incremental Parser), which was the main tool for NLP at Xerox for 20 years, and two different programming languages now in Open Source: Tamgu and LispE. 

Presentation

Claude kindly shared his presentation in case you missed his talk: Revolutionizing Agent Architecture with Logical Foundations

Abstract

Bayesian inference is a statistical learning method based on Bayes' theorem a representation of the changing beliefs. It simply demonstrates that the probability of a ‘hypothesis’ being correct becomes more reliable with supporting ‘evidence’. 

In this talk, Ritesh will present the application of the Bayesian inference method in the domain of natural hazard assessment and mitigation. Specifically, I've applied this method to a numerical model of a rockfall protection structure for two purposes.

First, to calibrate it against real-scale field test data and second to access its impact response through inverse analysis. The ease of implementing the inference methods is supported by a metamodeling technique.

Ritesh will also share an insight into the numerical model developed using the non-smooth contact dynamics (NSCD) method. This model presents three key characteristics favouring the implementation of a statistical learning method; (1) a reduced-order model, (2) about 20 times faster computational speed compared to the equivalent FE model and (3) a user-friendly interface for design and analysis. 

Short Bio

Ritesh as a post-doctoral researcher at INRAe Grenoble in 2022-23 where he worked on Bayesian inference and metamodelling technique application on rockfall protection structures numerical modelling.

He has previously worked at 3SR and Inria Grenoble on the uncertainty quantification (UQ) techniques for the durability assessment of nuclear containment building and for rockfall trajectory exploration respectively, and is keen to further explore new techniques in ''Data-Driven-Modelling'' for multiphysics processes and multi-fidelity data fusion.

Obtained his PhD from UGA in 2020, in the domain of Offshore Geotechnical Engineering, he is experienced in the finite element (FE) analysis and constitutive modelling of soil-structure interaction. 

Presentation

Ritesh kindly shared his presentation in case you missed his talk: Bayesian inference: Basics and application for an engineering structure