A persistent challenge in AI is the effective integration of plausible and formal reasoning - the former concerning the plausibility and contextual relevance of arguments, the latter focusing on their logical and structural validity. Large Language Models (LLMs) are not immune to such a challenge. By virtue of their extensive pre-training, LLMs can generate plausible and linguistically fluent arguments, but struggle with the systematicity and consistency required for robust logical reasoning. At the same time, LLMs offer new opportunities to study and overcome this intrinsic conflict. This talk will focus on such opportunities, presenting different research directions aimed at reconciling plausible and formal reasoning, including LLM-driven neuro-symbolic integration, quasi-symbolic abstractions, and latent circuit disentanglement. The final part of the talk will discuss the persisting challenges in achieving truly unified reasoning and outline possible directions for future research in the field.

Marco is a Lecturer in Artificial Intelligence and Applications of AI in the Natural Language Processing (NLP) group at the University of Sheffield. Prior to Sheffield, he was a member of the Neuro-Symbolic AI Group at the Idiap Research Institute in Switzerland, and obtained a PhD in Computer Science from the University of Manchester. His research focuses on developing AI systems that can use explanation as a core mechanism for learning and reasoning, investigating the integration of neural and symbolic AI methods. Moreover, he is interested in developing methodologies to interpret, control, and evaluate Large Language Models (LLMs), with a focus on disentangling knowledge acquisition from abstract logical reasoning, and enabling out-of-distribution, out-of-domain generalisation.

There is increasing interest in understanding multilingual training dynamics and shared representations, instead of analysing final model checkpoints. Tracing training dynamics allows us to analyse when linguistic information and shared concept spaces emerge during pre-training and understand model mechanisms, e.g. where alignment breaks down. In this talk, I will discuss why studying training dynamics is useful, particularly from a multilingual lens, and present recent findings on studying model behaviour and representations during pre-training.

Barbara Plank is Professor and co-director of the Center for Information and Language Processing at LMU Munich. She holds the Chair for AI and Computational Linguistics at LMU Munich and is a visiting Professor at the Computer Science department at the IT University of Copenhagen. Her MaiNLP research lab (Munich AI and NLP lab) focuses on robust machine learning for Natural Language Processing with an emphasis on human-inspired and data-centric approaches. Her research has been funded by distinguished grants and awards, including an ERC Consolidator Grant, DFF Sapere Aude Research Leader grant, ELLIS Fellow, and several best paper awards. She regularly serves on international committees, including the Association for Computational Linguistics (ACL), the European Chapter of the ACL, the Northern European Association for Language Technology (NEALT) and Scientific Advisory Boards of Research Centers across Europe.

I will argue that visual reasoning is a fundamental capability and one that has tremendous potential in multimodal language models. I will start by outlining the types of tasks that multimodal models still fall short on, drawing on decades of computer vision research. Next, I will introduce the concept of sketching, which operationalizes visual reasoning using external computer vision models as tools. I will demonstrate the potential of visual reasoning with sketching, and outline the limitations. After which, we will overcome these limitations by incorporating visual reasoning directly into the language model using perception tokens. Finally, I will describe how visual reasoning can enable robots to reason in space, allowing them to surpass non-reasoning proprietary robotics foundation models.

Ranjay Krishna is an Assistant Professor at the Allen School of Computer Science & Engineering. He co-directs the RAIVN lab at UW and directs the PRIOR team at the Allen Institute. His research lies at the intersection of computer vision, natural language processing, robotics, and human computer interaction. This research has received best paper honorable mentions at CVPR'25 and CSCW'23, outstanding paper at NeurIPS'21 and ACL'21, and dozens of orals at CVPR, ACL, CSCW, NeurIPS, UIST, and ECCV, and has been reported by Science, Forbes, the Wall Street Journal, and PBS NOVA. He is also recognized as one of MIT Technology Review's 35 under 35 Asia Pacific '25. His research has been supported by Google, Apple, Ai2, Amazon, Cisco, Toyota Motor Inc, Toyota Research Institute, NSF, ONR, and Yahoo. He holds a bachelor's degree in Electrical & Computer Engineering and in Computer Science from Cornell University, a master's degree in Computer Science from Stanford University and a Ph.D. in Computer Science from Stanford University.

Neural speech interfaces aim to restore natural communication in individuals who have lost the ability to speak while preserving cognitive function. Over the past decade, this field has undergone a remarkable transformation, moving from slow and cognitively demanding spelling-based brain–computer interfaces to systems capable of decoding continuous speech directly from neural activity. These advances have been driven by the convergence of high-resolution invasive neural recording technologies, improved experimental paradigms for speech production and perception, and powerful deep learning models inspired by modern automatic speech recognition systems. In this talk, I will review the state of the art in neural speech prostheses, with a particular focus on next-generation BCIs that translate cortical activity into text or synthetic speech. I will discuss key design choices, including neural recording techniques (such as ECoG, sEEG, and intracortical microelectrodes), target brain areas, decoding architectures, and evaluation metrics. I will also highlight recent clinical results demonstrating unprecedented levels of accuracy, fluency, and long-term stability in continuous speech decoding. Finally, I will outline current challenges and future directions, including scalability across users, real-time bidirectional feedback, and the path towards clinical and real-world deployment, illustrated with ongoing work from our research group.

Jose A. Gonzalez-Lopez is an Associate Professor at the University of Granada whose research sits at the frontier of artificial intelligence, computational neuroscience, and neural speech prostheses. His work addresses the core challenge of how to translate high-dimensional neural activity into fluent, natural speech, bridging invasive neural recordings with modern deep learning and speech–language models. He leads multiple competitive R&D projects on AI-driven speech restoration for individuals with severe neurological and phonatory impairments, with a strong emphasis on long-term robustness, scalability across users, and real-world clinical deployment. He has published over 100 papers in leading international journals and conferences. His contributions have been recognized with several awards for scientific excellence and technological innovation, and his research is embedded in a strong international collaboration network built through extended research visits to institutions such as the University of Sheffield, the University of Bremen, and Maastricht University.

Today's large language models (LLMs) are optimized toward giving helpful answers in response to prompts. In many situations, however, it may be preferable for an LLM to foster critical thinking rather than just following an instruction. While recent LLMs are said to 'reason', they barely build on established reasoning concepts known from argumentation theory. In this talk, I will give insights into recent efforts of my group in making LLMs more argumentative. Starting from basics of LLM training processes, I will present how to specialize LLMs for argumentation tasks via instruction fine-tuning as well as how to align the arguments they generate using reinforcement learning. From there, I will give an outlook on how to improve the actual reasoning capabilities of LLMs.

Henning Wachsmuth leads the Natural Language Processing Group at the Institute of Artificial Intelligence of Leibniz University Hannover. After receiving his PhD from Paderborn University in 2015, he worked as a PostDoc at Bauhaus-Universität Weimar and as a junior professor in Paderborn, before he became a full professor in Hannover in 2022. His group does basic research on large language models for computational argumentation, social bias detection and mitigation, as well as explainable and educational NLP. Henning's main research interests include the generation of audience-aware text, the assessment of pragmatic text quality, and the modeling of bias and framing.

Low resource NLP is nowadays an umbrella keyword that covers a wide set of research directions. In this talk I will argue that it is important to carefully present the low resource scenario and distinguish when the languages are truly lacking resources, as opposed to simulating lack of labeled data. I propose to follow a more systematic way to represent work in this space and to question how we approach technology development for these languages. I will also present recent work in my group that contributes to improve language representation by exploring efficient approaches to diverse languages.

Thamar Solorio is a professor of NLP at MBZUAI where she also serves as Vice Provost for Faculty Excellence and Advancement. Her research interests include NLP for low-resource settings and multilingual data, including code-switching and information extraction. More recently, she has been exploring language and vision problems, focusing on developing inclusive NLP. She served two terms as an elected board member of the North American Chapter of the Association of Computational Linguistics (NAACL) and was PC co-chair for NAACL 2019, and recently stepped down from being co-Editor in Chief of the ACL Rolling Review Initiative (ARR). She was the general chair of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP).

Language models tend to excel in languages they see the most during (pre)training—leaving low-resource languages at a stark disadvantage. But what if we could boost performance without throwing (much) more data or compute at the problem? In this talk, I’ll present a set of resource-lean (read: "cheap") strategies that enhance multilingual language understanding and generation in low-resource settings. I’ll show how conceptually effective knowledge transfer techniques—not just bigger models—can improve multilingual capabilities across three major fronts: (1) standard text-based LLMs, (2) vision-language models, and (3) code language models. The takeaway? Scaling isn’t the only answer: for truly inclusive multilingual language technology, we need stronger inductive biases and more conceptual innovation.

Goran Glavaš is a Full Professor for Natural Language Processing at the University of Würzburg (Germany), Center for AI and Data Science (CAIDAS). His research focuses on multilingual language understanding and cross-lingual transfer, vision-and-language models, and trustworthiness of (multilingual) language models. He has (co-)authored over 120 publications in NLP and IR, regularly publishing at top-tier venues (ACL, EMNLP, NAACL, EACL, TACL, SIGIR, ECIR). He received the best long paper award at EACL 2021 and outstanding paper awards at EACL 2024 and ACL 2024. He served as an Editor-in-Chief of the ACL Rolling Review (ARR) and regularly serves as (Senior) Area Chair for top-tier NLP conferences.

Despite recent tendencies towards building large "monolithic" neural models, fine-tuned expert models and parameter-efficient specialised modules still offer gains over large monoliths in specific tasks and for specific data distributions (e.g., low-resource languages or specialised domains). Moreover, such modularisation of skills and expertise into dedicated models or modules allows for asynchronous, decentralised, and more efficient continuous model development, as well as module reusability. However, a central question remains: how to combine and compose these modules to enable positive transfer, sample-efficient learning, and improved out-of-domain generalisation. In this talk, after discussing the key advantages of modularisation and modular specialisation, I will provide an overview of prominent module and model composition strategies. I will focus on composition at the parameter level (model merging) and functional level (model MoErging), and then illustrate the usefulness of these techniques across several applications.

Ivan Vulić is currently a Research Scientist at Google DeepMind in Zurich after spending a year there as a Visiting Researcher. Before that he was a Research Professor and a Royal Society University Research Fellow in the Language Technology Lab, University of Cambridge, where he spent 10 years across different research roles. From January 2018 until November 2024 he was also a Senior Scientist at PolyAI in London. Ivan holds a PhD in Computer Science from KU Leuven awarded summa cum laude. In 2021 he was awarded the annual Karen Spärck Jones Award from the British Computing Society for his research contributions to Natural Language Processing and Information Retrieval. His core expertise and research interests span, among others, cross-lingual, multilingual and multi-modal representation learning, modularity and composability of ML models, sample-efficient, parameter-efficient and few-shot ML, conversational AI, data-centric ML.

Language models have become ubiquitous in NLP and beyond. In particular, the new wave of large language models (LLMs) are increasingly used to communicate and solve practical problems in many languages and countries, and by an increasingly diverse set of users. However, even though there is no doubt that these models open up plenty of opportunities, there are important issues and research questions that arise when it comes to LLMs and their application in different languages and cultures. For instance, the language coverage in language models drastically decreases for less-resourced languages and as such, their performance. And not only the general performance is affected, but general-purpose LLMs may be implicitly biased to specific cultures and languages depending on their underlying training data. In this talk, I will discuss how language models reflect on cultural diversity, including potential shortcomings and how language coverage and cultural awareness may be intrinsically intertwined. I will also share some lessons learned based on our recent research in this area, including a large effort to develop a cultural benchmark of everyday knowledge for dozens of languages and countries.

Jose Camacho-Collados is a UKRI Future Leaders Fellow and Professor at the School of Computer Science of Cardiff University, where he co-founded the Cardiff Natural Language Processing group (Cardiff NLP). Before joining Cardiff University, he completed his PhD in Sapienza University of Rome and was a Google AI PhD Fellow. Jose has worked in multiple NLP areas with a particular focus on semantics, multilinguality and computational social science with an interdisciplinary perspective. In this area, he has been developing specialised and efficient NLP models for social media applications, such as TweetNLP and related efforts. His work has received several recognitions, including awards at top NLP conferences, and the 2023 AIJ Prominent Paper Award. He is also the co-author of the "Embeddings in Natural Language Processing" book.