UCLA NLP Seminar Series - Archive

Speaker Bio: Sean Welleck is an Assistant Professor at Carnegie Mellon University, where he leads the Machine Learning, Language, and Logic (L3) Lab. His areas of focus include generative models, algorithms for large language models, and AI for code, science, and mathematics. Sean received a Ph.D. from New York University. He was a postdoctoral scholar at the University of Washington and the Allen Institute for Artificial Intelligence. He is a recipient of a NeurIPS 2021 Outstanding Paper Award, and two NVIDIA AI Pioneering Research Awards.

Abstract: One of the most striking findings in modern research on large language models (LLMs) is that scaling up compute at training time leads to better final results. However, there is another lesser-mentioned scaling phenomenon, where adopting more sophisticated methods and/or scaling compute at inference time can result in significantly better outputs from LLMs. In this talk, I will talk about our lab's recent work on using inference-time strategies to enable better reasoning. This includes training models to think prior to steps of formal mathematical proving, leveraging strong evaluation models to enable easy-to-hard generalization, and inference scaling laws that optimally balance cost and performance. Together, these advances point to a new paradigm of scaling compute at inference time.

Speaker Bio: Sara Hooker leads Cohere For AI, the dedicated research arm of Cohere. Cohere For AI seeks to solve complex machine learning problems and supports fundamental research that explores the unknown. With a long track-record of impactful research at Google Brain, Sara brings a wealth of knowledge from across machine learning. Her work has focused on model efficiency training techniques and optimizing for models that fulfill multiple desired criteria -- interpretable, efficient, fair and robust. Sara leads a team of researchers and engineers working on making large language models more efficient, safe and grounded. Sara is currently on Kaggle's ML Advisory Research Board and serves on the World Economic Forum council on the Future of Artificial Intelligence.

Speaker Bio: Natasha Jaques is an Assistant Professor of Computer Science and Engineering at the University of Washington, and a Senior Research Scientist at Google DeepMind. Her research focuses on Social Reinforcement Learning in multi-agent and human-AI interactions. During her PhD at MIT, she developed techniques for learning from human feedback signals to train language models which were later built on by OpenAI’s series of work on Reinforcement Learning from Human Feedback (RLHF). In the multi-agent space, she has developed techniques for improving coordination through social influence, and unsupervised environment design. Natasha’s work has received various awards, including Best Demo at NeurIPS, an honourable mention for Best Paper at ICML, and the Outstanding PhD Dissertation Award from the Association for the Advancement of Affective Computing. Her work has been featured in Science Magazine, MIT Technology Review, Quartz, IEEE Spectrum, Boston Magazine, and on CBC radio, among others. Natasha earned her Masters degree from the University of British Columbia, undergraduate degrees in Computer Science and Psychology from the University of Regina, and was a postdoctoral fellow at UC Berkeley.

Abstract: If AGI is right around the corner, why are AI agents still so bad at so many tasks? AI still fails to coordinate effectively with other agents, follow natural language instructions to complete embodied tasks, and generalize to circumstances not encountered during training. Even in pure language settings like dialog, AI still fails to adapt to the needs of individual users, instead aligning to a single set of values that may ignore the needs of minority groups. In this talk, I will argue that Social Learning is a key facet of intelligence that enables both humans and animals to easily adapt to new circumstances, coordinate with different people, and acquire complex behaviors. By improving the social intelligence of AI agents, we can get a step closer to adaptive, flexible, generalist agents which better align to diverse human values. This talk will overview recent work in the Social Reinforcement Learning lab, describing how to enable pluralistic alignment of large language models using human feedback, smooth coordination with diverse human partners, and improve social reasoning for understanding natural language commands.

Speaker Bio: Alexander Spangher is pursuing his PhD in computer science at the University of Southern California; he is formerly a writer and data scientist at the New York Times. He focuses on computational journalism and is advised by Jonathan May, Emilio Ferrara and Nanyun Peng. His research is broad and has pursued the following side directions: he has worked at Microsoft Research under the mentorship of Eric Horvitz to detect misinformation. He has collaborated with EleutherAI to build state-of-the-art symbolic music models. Finally, he has collaborated with the MIT Plasma Science and Fusion Center (PFSC) to model disruptions in nuclear fusion reactions. His work has received numerous awards: 2 Outstanding Paper Awards at EMNLP 2024, 1 Spotlight Award at ICML 2024, and an Outstanding Paper Award at NAACL 2022. He is fortunate to be supported by a 4-year Bloomberg PhD Fellowship.

Abstract: Recent modeling innovations incorporate planning — or reasoning about actions (exhibited by models like GPT-o1 and Deepseek's R1) — and have demonstrated impressive performance gains in areas like mathematical problem-solving and computer coding. However, such domains are characterized by well-defined goals (or rewards). For many human-centered tasks in creative contexts, rewards are not as clearly defined and it is thus not clear how to make similar progress in these domains. In this talk, I will outline a research agenda that can enable us to make progress in these fundamentally human processes. I focus on tasks related to journalism, where there is a pressing need for technical innovation. Specifically, in this talk I will focus on the task of retrieving sources relevant to news stories: I will show how (1) we can make inferences about human actions based on environmental state-observations (a process known to cognitive psychologists as "end-state" or "ghost conditions", but as yet unexplored in machine learning); and, (2) how these inferences can help us learn human values and rewards.

Speaker Bio: I am a Researcher at Anthropic. I am primarily interested in reasoning, AI for science and AI alignment. Previously, I worked on reasoning and problem solving in language models at OpenAI. I contributed to the recent OpenAI o1 models, a new state-of-the-art in LLM reasoning. I have also worked on weak-to-strong-generalization on the superalignment team under Jeff Wu, Jan Leike and Ilya Sutskever. I also had a short stint at xAI, where I reported to Elon Musk. Starting in Fall 2025, I will be joining NYU as an Assistant Professor in the Tandon CSE department, and Courant CS department by courtesy. I am also a member of the NYU CILVR Group. I defended my PhD in Computer Science at NYU, in 2023.

Abstract: Widely used alignment techniques, such as reinforcement learning from human feedback (RLHF), rely on the ability of humans to supervise model behavior—for example, to evaluate whether a model faithfully followed instructions or generated safe outputs. However, future superhuman models will behave in complex ways too difficult for humans to reliably evaluate; humans will only be able to weakly supervise superhuman models. We study an analogy to this problem: can weak model supervision elicit the full capabilities of a much stronger model? We test this using a range of pretrained language models in the GPT-4 family on natural language processing (NLP), chess, and reward modeling tasks. We find that when we naively finetune strong pretrained models on labels generated by a weak model, they consistently perform better than their weak supervisors, a phenomenon we call weak-to-strong generalization. However, we are still far from recovering the full capabilities of strong models with naive finetuning alone, suggesting that techniques like RLHF may scale poorly to superhuman models without further work. We find that simple methods can often significantly improve weak-to-strong generalization: for example, when finetuning GPT-4 with a GPT-2-level supervisor and an auxiliary confidence loss, we can recover close to GPT-3.5-level performance on NLP tasks. Our results suggest that it is feasible to make empirical progress today on a fundamental challenge of aligning superhuman models.

Speaker Bio: Faeze Brahman is a Research Scientist at the Allen Institute for AI (Ai2). Prior to that, she was a postdoctoral researcher at Ai2 and the University of Washington and received her PhD from UCSC. Her research focuses on constrained reasoning and generation, understanding LLMs' capabilities and limitations and bridging the capability gap between humans and models beyond scaling through developing resource-efficient algorithms. She is also interested in designing human-centered AI systems that are reliable and safe for real-world applications.

Abstract: As AI assistants grow increasingly capable, their responsible deployment relies not just on what they can do, but on knowing when not to comply. Moving beyond traditional safety-focused view of AI noncompliance, I will talk about two projects that tackle this challenge: First, I introduce a taxonomy of contextual noncompliance, that identifies when and how models should handle misleading, out-of-scope or underspecified requests—revealing significant gaps in current systems' ability to do so. Second, I present a selective evaluation framework that enables models to abstain from making unsound judgments when they lack confidence, achieving stronger alignment with human evaluators while remaining cost-effective. Together, these works help create AI systems that are more reliable and safe to use across diverse real-world use cases.

Speaker Bio: Robin Jia is an Assistant Professor of Computer Science at the University of Southern California. He received his Ph.D. in Computer Science from Stanford University, where he was advised by Percy Liang. He has also spent time as a visiting researcher at Facebook AI Research, working with Luke Zettlemoyer and Douwe Kiela. He is interested broadly in natural language processing and machine learning, with a focus on scientifically understanding NLP models in order to improve their reliability. Robin’s work has received best paper awards at ACL and EMNLP.

Abstract: The rise of large language models offers opportunities to both scientifically study these complex systems and apply them in novel ways. In this talk, I will describe my group’s recent work along these lines. First, I will discuss data watermarks, a statistically rigorous technique for auditing a language model’s training data based only on black-box model queries. Then, we will investigate how language models memorize training data: based on results from two complementary benchmarks, I will demonstrate the viability of localizing memorized data to a sparse subset of neurons. Next, I will provide a mechanistic account of how pre-trained language models use Fourier features to solve arithmetic problems, and how pre-training plays a critical role in these mechanisms. Finally, I will show how to leverage the complementary strengths of large language models and symbolic solvers to handle complex planning tasks.

Speaker Bio: Jieyu Zhao is an assistant professor of Computer Science Department at University of Southern California where she is leading the LIME lab. Prior to that, she was an NSF Computing Innovation Fellow at University of Maryland, College Park. Jieyu received her Ph.D. from Computer Science Department, UCLA. Her research interest lies in fairness of ML/NLP models. Her research has been covered by news media such as Wires, The Daily Mail and so on. She was invited by UN-WOMEN Beijing on a panel discussion about gender equality and social responsibility.

Abstract: The rapid advancement of large language models (LLMs) has unlocked a myriad of possibilities for positive societal impact, ranging from enhancing accessibility and communication to supporting disaster response and public health initiatives. However, the deployment of these technologies also raises critical concerns regarding accountability, fairness, transparency, and ethical use. In this talk, I will discuss our efforts for auditing NLP models, detecting and mitigating biases, and understanding how LLMs make decisions. We hope to open the conversation to foster a community-wide effort towards more accountable and inclusive NLP practices.

Speaker Bio: Elisa Kreiss is an Assistant Professor of Communication at UCLA and the lab director of the Coalas (Computation and Language for Society) Lab. Previously, she completed a PhD in Linguistics at Stanford, where she was a member of Stanford’s NLP group and the Stanford Data Science Center for Open and REproducible Science (CORES). Elisa investigates how we produce and understand language situated in the visual world. Her work combines tools from natural language processing, psycholinguistics, and human-computer interaction to advance our understanding of how communicative context shapes language use. Her research has direct applications to image accessibility – the challenge of (automatically) generating image descriptions for blind and low-vision users. Elisa’s work has been supported by several Google Research Awards, the National Science Foundation, Stanford’s Human-centered AI initiative, and Stanford’s Accelerator for Learning.

Abstract: Developing Vision-Language Models (VLMs) that can easily translate between the linguistic and visual modality in human-like ways has many useful applications, including making visual content accessible to blind and low vision individuals, detecting misinformation, and combating visual illiteracy. While the current generation of VLMs has quickly risen to show human-level performance on many existing benchmarks, there remains a remarkable gap between these scores and how useful the models are found to be in practice. In this talk, I will present recent and ongoing work which suggests that in order to develop and understand the merit of Vision-Language Models for downstream application, we need to define tasks and evaluation metrics that assess the communicative usefulness of the generated texts. Specifically, I will focus on the challenge of generating image descriptions and argue for moving the goal post from what can be said about an image to the fundamentally pragmatic question of what should be said about it. Based on a variety of experiments with sighted and blind and low-vision participants, I will show that the pragmatic notion of contextual relevance is a core pillar of generating human-like image descriptions, provide evidence that our current tasks and evaluation tools in NLP remain unhelpful in uncovering these context effects, and present work that starts addressing this gap. Taken together, this work provides fundamental insights into how people communicate about the visual world, and shows how we can use those insights to advance VLMs for social impact, such as non-visual accessibility.