Dzmitry Bahdanau (ServiceNow Research). Agents and Reasoning
[Abstract]
This lecture delves into two dynamic areas of AI research and development: agents and reasoning. The idea behind AI Agents is to give neural AI human-like agency to take critical decisions. We will cover evidence in favor and against this radical approach, as well as recent evaluations of long-term agentic autonomy. A common belief in the AI community is that agents can become more autonomous and reliable if they are allowed to perform more computation to choose the next action. We will discuss the compute-heavy reasoning approaches recently demonstrated by OpenAI and DeepSeek and review the evidence of their relevance for agentic autonomy.
Alexandra Birch (University of Edinburgh). Reasoning and Knowledge in Multilingual Models
[Abstract]
Large language models are mainly trained on English text leading to lower performance in other languages. In this talk we will look at analysing the way that LLMs process different tasks across languages, highlighting the importance of neurons which are shared across languages. We then look a number of ways of improving LLMs multilingual performance in reasoning and knowledge intensive tasks.
Kyunghyun Cho (New York University & Prescient Design)
Marzieh Fadaee (Cohere For AI). Languages as Windows to Human Knowledge: Building Better Language Models Through Global Understanding
[Abstract]
Modern language models have achieved remarkable capabilities in English, but human knowledge and experience span thousands of languages, each encoding unique perspectives and problem-solving approaches. From the algorithmic precision required to handle Arabic’s root-pattern morphology to the contextual reasoning needed for Japanese’s topic-prominent structure, each language presents distinct computational challenges that push the boundaries of natural language processing.
The talk will address pressing challenges in multilingual AI development, including training data imbalances, cross-lingual transfer limitations, safe and harmless generations, and evaluation complexity across different languages. I’ll discuss practical solutions and emerging research directions that could help bridge the current performance gap between high-resource and low-resource languages. By building truly multilingual AI systems, we not only expand technology access but also develop more sophisticated models capable of handling the full spectrum of human language complexity.
Benoît Favre (Aix-Marseille University, LIS, CNRS, LIG). Reaching out to Other Modalities in Language Modeling
[Abstract]
Recent advances in natural language processing have allowed
very impressive technologies connecting language with other modalities
such as audio or images. This includes applications such as translating
from one modality to the other (speech to text, image captioning, but
also text to speech or image generation from a description), textual
interaction with other modalities (visual question answering, acoustic
event qualification), and even real-time spoken dialogue with an agent.
These new capabilities are expected to fuel the upcoming generation of
robots in order to give them more human-like capabilities, and a more
natural interface through language. Audio and image modalities are
typically integrated in a LLM by first transforming non-textual inputs
as a set of vectors, and projecting these representations to the same
spaces as text. Models can then be trained to perform tasks such as
responding to instructions on mixed modality inputs. In this talk, I
will give an overview of past and current approaches for handling
modalities in language models. In particular, I will focus on foreign-modality tokenization techniques and the properties of the underlying
representation space.
Dirk Hovy (Bocconi University). From Words to Culture: Unleashing NLP for Social Science Insights
[Abstract]
The rapid development of large language models in recent years has transformed the field of NLP. Many people are concerned that it has trivialized the field or even rendered it obsolete.
In this talk, I’ll argue that neither is true: NLP has a long way to go, and LLMs are the most recent in a long line of methods that have advanced the field. LLMs have freed us from many of the nitty-gritty details that previously hampered NLP research, allowing us to focus on larger and more interesting questions.
Computational social science is one area that still has many interesting questions and research projects to pursue. There is a lot that NLP models can tell us about language, and that language can tell us about society. From relatively simple text classification approaches for sentiment and social media analysis to capture political shifts and company culture; latent structure discovery through clustering-related approaches for socio-linguistic developments; hidden societal biases about religion, gender, socio-economic status, and emotions; the ability to reframe political arguments to address content moderation; prototyping survey studies, inferring missing demographic groups, and deepening the response validity; to comparing human and algorithmic preferences and what that tells us about being human.
This talk will provide a comprehensive practical guide for aspiring computational social scientists interested in using NLP to work with text data. We will also discuss the limitations of LLMs regarding social interaction, the implications for their safety and utility, and what this means for the future of NLP research.
François Yvon (CNRS, ISIR). Text Generation: Know Your Options !
[Abstract]
Text generation, contextual or non-contextual, is ubiquitous in the
current LLM era, as it serves as the most basic block in multiple
application contexts, from question answering and dialog systems to text
summarization and machine translation, and many more. Generation is thus
equally useful to compute deterministic and highly non-deterministic
mappings with various level of output constraints. Furthermore, text
generation is also used as a sub-routine of more complex generation
strategies, aiming to produce syntactically well-formed (e.g. for code
generation) or semantically consistent outputs, possibility through
multiple steps of generation (e.g, in chain-of-thoughts generation) or
to collect diverse samples from the generating distribution. To cover
this considerable diversity of uses, multiple text generation strategies
have been proposed, some less well-known than others. In this talk I
will review various families of generation algorithms, from the most
basic ones to the more sophisticated approaches, so as to document, as
much as possible, the possible options that are available to text
generation users.
Speakers from previous editions are here (2024), here (2023), here (2022) and here (2021)