Adaptive Memory-Augmented Agentic Reasoning (2026-2028 Huawei)
Adaptive Memory-Augmented Agentic Reasoning (2026-2028 Huawei)
We advance towards Adaptive Agentic Reasoning systems in which parametric knowledge is continuously complemented by dynamic memory and are not passive repositories of knowledge but active reasoners: they can plan retrieval, orchestrate tools, update and query memory, and adapt through interaction and oversight mechanisms
We are creating a multilingual reasoning benchmark to be able to evaluate frontier models on their ability to perfrom long-horizon tasks in multiple languages. The benchmarks will allow for fine-grained evaluation of LLM abilities to complete tasks with checkpoints for tasks and both monolingual and cross-lingual abilities tested. We will also test safety and culturally specific compliance of LLMs.
The creation of a wiki and a whitepaper providing a comprehensive and practical guide to best practices, designed to help those in research and in industry to uphold the highest ethical standards in their work. Wiki and Whitepaper
An Extreme Scale EuroHPC grant on Jupiter with the goal of creating a Suite of Multimodal Reasoning Foundation Models for Europe following on from EuroLLM
A 1.7M GPU hour grant from EuroHPC on the Barcelona Super Computer MareNostrum5 to train a language model for Europe. It covers the 24 official European languages and 11 strategic international languages. We pretrained a 1.7B and 9B model from scratch on 50% non-English data. EuroLLM shows the best translation performance compared to all the other top medium sized LLMs. It also performs competitively in understanding and reasoning across 11 EU languages, beating other European models, and it is fully open.
Investigate the Full-Duplex (FD) capability of Large Language Models (LLMs). Unlike half-duplex models, FD models [2] can listen and speak simultaneously. This property is particularly important in natural human conversations, where participants often backchannel (e.g., “uh-huh”), overlap, or interrupt each other. We conducted a series of reproducible and comparable experiments to tailor a recipe to equip LLMs the capability to conduct FD conversations using academic scale of data and compute. Report
MTStretch aimed to improve low-resource machine translation by stretching our capabilities and resources.
MTStretch was a EPSRC Innovation Fellowship which ran from 2019-2021.
Global Under-Resourced Media Translation
GoURMET improves neural machine translation for low-resource language pairs and domains.
The project develops open-source applications and implements these through use cases at the BBC and Deutsche Welle.
GoURMET is a European Horizon 2020 project which ran from 2019-2022
Scalable Understanding of Multilingual Media
SUMMA developed an extensible media monitoring platform and tools, to enhancing Multilingual and cross lingual capabilities. It was an EU Horizon 2020 project that ran from 2016-2019.
An industrial collaboration with Samsung Research Poland, covering a broad range of challenges for neural machine translation lasting from 2015-2020.
Scalable Understanding of Multilingual Media
SUMMA developed an extensible media monitoring platform and tools, to enhancing Multilingual and cross lingual capabilities. It was an EU Horizon 2020 project that ran from 2016-2019.
Health in my language.
Making public health information available in a wider variety of languages. It was an EU Horizon 2020 project which ran from 2015-2018