Talk Title: When Best-of-N is Worse: Coverage, Reward Hacking, and Pessimism in Inference-Time Alignment

Abstract: Inference-time scaling has emerged as a powerful way to improve language models, but its simplest form, Best-of-N (BoN) sampling, has a delicate failure mode.  While additional samples can substantially improve outputs, in the presence of imperfect verifiers, BoN can amplify reward-hacking by over-optimizing to the learned reward model.  In this talk, we will discuss a perspective on inference-time alignment centered on this tension.  I will argue that the performance of BoN is governed by three key ingredients: the coverage of the reference model, the quality of the reward model, and the metric used to measure success.  We will see that, under appropriate coverage assumptions, BoN can be optimal for improving expected reward, but that these guarantees do not by themselves rule out sever reward hacking or poor win-rate performance.  This helps explain why seemingly contradictory conclusions can all be correct, depending on what objective is being optimized and how inference-time success is measured.  Motivated by this, I will then discuss pessimistic alternatives to standard BoN, including methods that use uncertainty estimates at inference time to penalize suspicious high-reward responses.  These approaches retain the benefits of additional inference-time compute while mitigating over-optimization to flawed verifiers.  Overall, the talk will argue that making language models better at inference time is not just a question of scaling up search, but of understanding when extra compute is pointed in the wrong direction.

Bio: Adam Block is an Assistant Professor of Computer Science at Columbia University and an affiliate faculty member in the Department of Electrical Engineering. His research spans the theoretical foundations of machine learning, with a focus on designing algorithms that are both provably efficient and practically effective in sequential decision-making and reinforcement learning settings, including work relevant to aligning learning systems with complex objectives. He completed his Ph.D. in Applied Mathematics and Statistics at the Massachusetts Institute of Technology, where he was affiliated with the Laboratory for Information and Decisions Systems and the Statistics and Data Science Center, and his dissertation focused on theory and applications of online learning. Prior to joining Columbia, he was a postdoctoral scientist at Microsoft Research New York City. Block received his B.A. in Mathematics, summa cum laude, from Columbia University and his work bridges core theoretical insights with challenges in modern machine learning and interactive learning environments.

Talk Title: The Creative Limits of Next-token Prediction

Abstract: Current LLM pipelines rely on a convenient illusion: that scaling next-token prediction and tweaking temperature naturally unlocks diverse, open-ended generation. In reality, standard autoregression is fundamentally myopic. We quantify this using minimal algorithmic tasks that require far-sighted stochastic planning. In these environments, next-token learning fails to plan, whereas multi-token approaches excel. Furthermore, standard output-layer temperature sampling degrades coherence in its attempt to elicit randomness. Surprisingly, simply injecting noise directly at the input layer (seed-conditioning) works as well, if not better. This same diversity collapse plagues test-time compute scaling for math reasoning, where standard decoding merely regurgitates redundant errors. Yet, applying a simple mode-conditioning (ModC) prefix forces the model to explore distinct reasoning paths, instantly yielding a 4x efficiency gain.

Bio: Aditi Raghunathan is an Assistant Professor at Carnegie Mellon University. She works broadly in machine learning, and her goal is to make machine learning more reliable and robust. Raghunathan's work spans both theory and practice, and leverages tools and concepts from statistics, convex optimization, and algorithms to improve the robustness of modern systems based on deep learning.Until recently, Raghunathan was a postdoc at Berkeley AI Research. She received her Ph.D. from Stanford University in 2021 where she was advised by Percy Liang. Her thesis won the Arthur Samuel Best Thesis Award at Stanford. Previously, she obtained her BTech in Computer Science from IIT Madras in 2016.

Talk Title: Tree-vial Pursuits: How Humble Decision Trees Still Outsmart the Generative Giants

Abstract: Large Language Models have redefined our expectations for what AI can achieve, showing remarkable prowess in natural language, complex reasoning, and code synthesis. Given these leaps, it is tempting to assume that numerical fluency would follow as a natural byproduct of scale. However, the reality is far more humbling: even the most sophisticated LLMs often fail spectacularly at basic tabular prediction tasks. This gap is a significant bottleneck, considering that the vast majority of the world’s enterprise and scientific data remains locked in rows and columns. In this talk we investigate the dissonance between a model's linguistic confidence and its actual predictive performance and we explore where our false perception of the LLMs' numerical mastery could stem from.

Bio: Marta Garnelo leads the research team at Fundamental as Chief Science Officer, where she is currently focused on the challenge of building foundation models for tabular data. Before moving to Barcelona for this, she spent over seven years as a researcher at DeepMind, focusing on the intersection of probabilistic modeling and deep learning. Marta is perhaps best known for introducing Neural Processes, though her research spans a broad range of topics including meta-learning, multi-agent RL, game theory, and exploring alternatives to the attention mechanism. She earned her PhD from Imperial College London.

Talk Title: Towards Truly Open, Language-Specific, Safe, Factual, and Specialized Large Language Models

Abstract: As large language models increasingly shape knowledge, communication, and creativity, it is imperative that we make them open, language-specific, safe, factual, and specialized. First, we will argue for the need for fully transparent open-source large language models (LLMs), and we will describe the efforts of MBZUAI's Institute on Foundation Models (IFM) towards that based on the LLM360 initiative. Second, we will argue for the need for language-specific LLMs, and we will share our experience from building Jais, the world's leading open Arabic-centric foundation and instruction-tuned large language model, Nanda, our open-weights Hindi LLM, Sherkala, our open-weights Kazakh LLM, and some other models. Third, we will argue for the need for safe LLMs, and we will present Do-Not-Answer, a dataset for evaluating the guardrails of LLMs, which is at the core of the safety mechanisms of our LLMs. Fourth, we will argue for the need for factual LLMs, we will discuss the factuality challenges that LLMs pose. We will then present some recent relevant tools for addressing these challenges developed at MBZUAI: (i) OpenFactCheck, a framework for fact-checking LLM output, for building customized fact-checking systems, and for benchmarking LLMs for factuality, (ii) LM-Polygraph, a tool for predicting an LLM's uncertainty in its output using cheap and fast uncertainty quantification techniques, and (iii) LLM-DetectAIve, a tool for machine-generated text detection. Finally, we will argue for the need for specialized models, and we will present some other LLMs currently being developed at MBZUAI's IFM.

Bio: Preslav Nakov is Professor and Department Chair for NLP at Mohamed bin Zayed University of Artificial Intelligence. He leads the development of Jais, the world's best open-source Arabic-centric LLM, Nanda, the world's best open-weights Hindi model, and Sherkala, the world's best open-weights Kazakh model at MBZUAI's Institute of Foundation Models. Previously, he was Principal Scientist at the Qatar Computing Research Institute, HBKU, where he led the Tanbih mega-project, which aims to limit the impact of "fake news", propaganda and media bias by making users aware of what they are reading, thus promoting media literacy and critical thinking. He received his PhD degree in Computer Science from the University of California at Berkeley, supported by a Fulbright grant. He is Chair of the European Chapter of the Association for Computational Linguistics (EACL), Secretary of ACL SIGSLAV, and Secretary of the Truth and Trust Online board of trustees. Formerly, he was PC chair of ACL 2022, and President of ACL SIGLEX. He is also a member of the editorial board of several journals including Computational Linguistics, TACL, ACM TOIS, IEEE TASL, IEEE TAC, CS&L, NLE, AI Communications, and Frontiers in AI. He authored a Morgan & Claypool book on Semantic Relations between Nominals, two books on computer algorithms, and 250+ research papers. He received a Best Paper Award at ACM WebSci'2022, a Best Long Paper Award at CIKM'2020, a Best Resource Paper Award at EACL'2024, a Best Demo Paper Award (Honorable Mention) at ACL'2020, a Best Task Paper Award (Honorable Mention) at SemEval'2020, a Best Poster Award at SocInfo'2019, and the Young Researcher Award at RANLP’2011. He was also the first to receive the Bulgarian President's John Atanasoff award, named after the inventor of the first automatic electronic digital computer. His research was featured by over 100 news outlets, including Reuters, Forbes, Financial Times, CNN, Boston Globe, Aljazeera, DefenseOne, Business Insider, MIT Technology Review, Science Daily, Popular Science, Fast Company, The Register, WIRED, and Engadget, among others. 

Bio: Samy Bengio (PhD in computer science, University of Montreal, 1993) is a senior director of machine learning research at Apple since 2021. Before that, he was a distinguished scientist at Google Research since 2007 where he was heading part of the Google Brain team, and at IDIAP in the early 2000s where he co-wrote the well-known open-source Torch machine learning library.

His research interests span many areas of machine learning such as deep architectures, representation learning, vision and language processing and more recently, reasoning.

He is action editor of the Journal of Machine Learning Research and on the board of the NeurIPS foundation. He was on the editorial board of the Machine Learning Journal, has been program chair (2017) and general chair (2018) of NeurIPS, program chair of ICLR (2015, 2016), general chair of BayLearn (2012-2015), MLMI (2004-2006), as well as NNSP (2002), and on the program committee of several international conferences such as NeurIPS, ICML, ICLR, ECML and IJCAI.

Talk Title: Are Mixture-of-Experts Modular? Why It Matters and How to Fix It

Abstract: Mixture-of-Experts (MoEs) are designed as modular architectures—but are they functionally modular, i.e., enabling the independent use of expert subsets for downstream domains? We argue they are not, and that this gap matters: as MoEs grow larger, sparser, and more fine-grained, they become increasingly difficult to use, adapt, and fine-tune without heavy infrastructure. We introduce ModMoE, a self-supervised approach that makes modularity a first-class property—without human priors or loss in overall performance. ModMoE induces semantically specialized experts (rather than lexical partitioning) and enables effective selective expert usage across pool sizes, improving efficiency and performance in both zero-shot inference and fine-tuning. These results point toward more accessible and flexible MoEs, and a path to large-scale, sparse, and truly modular expert architectures.

Bio: Sewon Min is an Assistant Professor in EECS at UC Berkeley, affiliated with Berkeley AI Research (BAIR), and a Research Scientist at the Allen Institute for AI. Her research lies at the intersection of natural language processing and machine learning, with a focus on large language models (LLMs). She studies the science of LLMs and develops new models and training methods for better performance, flexibility, and adaptability, such as retrieval-based LMs, mixture-of-experts, and modular systems. She also studies LLMs for information-seeking, factuality, privacy, and mathematical reasoning. She has organized tutorials and workshops at major conferences (ACL, EMNLP, NAACL, NeurIPS, ICLR), served as a Senior Area Chair, and received honors including best paper and dissertation awards (including ACM Dissertation Award Runner-up), a J.P. Morgan Fellowship, and EECS Rising Stars. She earned her Ph.D. from the University of Washington and has held research roles at Meta AI, Google, and Salesforce.