Foundation-model agents are being deployed faster than the community can characterize how they fail. In barely two years they have moved into browsing, software engineering, computer use, scientific discovery, and embodied control, where reliability is governed by long-horizon, closed-loop interaction that current single-turn evaluation paradigms rarely capture. Failures (error cascades, reward hacking, mode and reasoning collapse, memory drift, and over-reasoning loops) still lack a shared vocabulary, reproducible triggers, and comparable diagnostics. This cutting-edge tutorial gives a systematic, mechanism-first account of agentic failure modes across four axes: (1) environments and infrastructure, (2) training dynamics, (3) evaluation and red teaming, and (4) test-time scaling and adaptation, with safety threaded throughout. Attendees leave with reproducible triggers, comparable diagnostic protocols, and verifiable mitigations.

Modern language models are not only tools for processing text; they are also becoming computational instruments for studying how humans understand, predict, and generate language. This tutorial introduces neural encoding and neural decoding to the NLP community through this broader perspective: how can NLP help us understand language processing in the human brain, and how can brain data help us evaluate and improve language models? We will present neural encoding as a way to test whether representations from language models—such as contextual embeddings, surprisal, attention patterns, and layer-wise features—can explain brain responses during reading or listening. We will then introduce neural decoding as the reverse problem: reconstructing linguistic or multimodal content from brain activity using retrieval, representation learning, and foundation-model-guided generation. Rather than framing decoding mainly as brain–computer interface technology, we will connect it to core NLP questions about meaning representation, generation under uncertainty, multimodal grounding, and semantic evaluation. The tutorial assumes no neuroscience background. It combines intuitive foundations, representative papers, practical modeling recipes, notebook-style demos, and discussion of mental privacy, consent, and responsible brain–language technologies.

Tokenization defines how data is represented as input and output for machine learning systems, including large language models (LLMs). While tokenization has long been a foundational component of NLP systems, it has historically remained a niche research area, often considered solved. The recent proliferation of LLMs has changed this: as these models are deployed across diverse languages and domains, tokenization-related issues have become increasingly apparent—from cross-lingual representation disparities to efficiency bottlenecks. This tutorial will provide a comprehensive overview of tokenization methods in text, starting from classical word-based and morphological approaches, to the current standard of statistical subword tokenizers, and emerging methods on end-to-end learnable tokenizers. We will also discuss the implications of tokenization choices on model performance, fairness, and efficiency, and present best practices for evaluating and designing tokenizers in multilingual and domain-specific settings.

Recent advances in large language models, formal reasoning, theorem proving, and agentic AI are rapidly changing what can be augmented and automated in mathematics. Yet mathematical reasoning is considerably richer than producing correct final answers: it involves understanding mathematical discourse, constructing and revising arguments, moving between informal and formal representations, exploring alternative proof strategies, verifying intermediate claims, and ultimately supporting mathematical discovery. This tutorial provides a structured introduction to the emerging field of automated mathematical reasoning, with a particular focus on autoformalization and neuro-symbolic and agentic approaches. We first examine mathematical reasoning from linguistic and conceptual perspectives, decomposing the spectrum from informal mathematical discourse to fully formal proofs. We then introduce the principal computational mechanisms underlying modern systems, including formal representations, autoformalization, theorem provers, critics and proxy evaluators, refinement loops, and complex agentic architectures. The tutorial connects these foundations with practical system-building, evaluation methodologies, current state-of-the-art results, and demonstrations of existing frameworks. We further examine how mathematical reasoning methods can extend toward physics, where mathematical arguments must interact with physical models, assumptions, approximations, and scientific knowledge. Finally, we use ambitious mathematical and scientific problems to identify the capabilities still required to move from benchmark solving toward systems that can participate meaningfully in verification, reasoning, and discovery.

Interpretability findings on large language models often fail to generalise, and causal claims can outrun the evidence. A recurring source, we suggest, is a borrowed vocabulary: interpretability inherited the words of causality — "mechanisms", "circuits", "features" — without always inheriting the bookkeeping that keeps causal claims precise. This tutorial teaches that bookkeeping for practitioners. Participants leave able to take any interpretability result and answer three questions: what claim does this method license, what would strengthen the claim, and when will it generalise?

While large language models (LLMs) have substantially expanded support for non-English languages, their ability to generalize across typologically diverse languages and produce contextually appropriate, culturally grounded responses remains a significant challenge. This tutorial examines the rapidly evolving area of understanding, evaluating, and improving the multilingual capabilities and multicultural alignment of LLMs. We will primarily focus on cutting-edge advances and latest publications from 2025–2026, while incorporating selected introductory material to provide necessary background and to ensure accessibility for a broad audience. Participants will gain an up-to-date view of both methods and practical insights, including multilingual pre-, mid- and post-training, multilingual-specific optimization techniques, culturally aware evaluation, and key open problems in developing more fair, inclusive, and robust LLMs.