..................................................................................4/14/2025..................................................................................

Title: LLM Agents — A Journey of Automation

Abstract: Creating intelligent agents capable of performing autonomous actions has long been a core objective of AI research. This talk highlights my journey in building systems that harness the knowledge and reasoning capabilities of large language models (LLMs) to achieve greater efficiency and autonomy across diverse environments, with a focus on games. The talk concludes with thoughts on the implications of our learnings for inference-time scaling.

..................................................................................3/19/2025..................................................................................

Title: Instruction Tuning: The Road to Intelligent LLMs 

Abstract: Large language models (LLMs) have demonstrated strong performance across numerous NLP tasks, yet they struggle to interact seamlessly with human users. Instruction tuning addresses this gap by training LLMs to understand and follow natural language instructions, which helps them evolve into intelligent AI assistants. In this lecture, I will first introduce the technical motivations behind instruction tuning. Next, I will explore how the instruction-following ability is evaluated and highlight some state-of-the-art techniques used in instruction tuning LLMs. Finally, I will discuss a few promising research directions aimed at developing next-generation instruction-following models.

..................................................................................2/21/2025..................................................................................

Title: Multi-Teacher Distillation: An Ensemble-Then-Distill Approach 

Abstract: Knowledge distillation (KD) aims to transfer the knowledge in a large model (called a teacher) into a small one (called a student), and has become an emerging research topic as the sizes of deep learning models keep growing. Today, there are abundant readily available large models, such as ChatGPT, LLaMa, and T5. It then becomes natural to ask: Can we distill the knowledge from multiple teachers? At first glance, it appears easy to perform multi-teacher KD, as we can simply train the student from the union of teachers’ predictions. However, I would argue that such a naïve attempt may not work well for multi-teacher KD. This is because traditional KD adopts the cross-entropy loss, which tends to yield a smooth distribution. In this talk, I will present a novel ensemble-then-distill approach, which builds an ensemble of teacher models to train the student. I will also discuss applications to text generation and syntactic parsing.

..................................................................................1/24/2025..................................................................................

Title: From Risk to Resilience: Navigating Misalignment in Large Language Models

Abstract: As large language models (LLMs) become central to intelligent systems, their use is expanding from everyday applications to high-stakes domains. Alignment plays a crucial role in the success of LLMs, ensuring that model behavior matches our expectations and remains consistent with various objectives. However, misalignment remains a significant challenge that undermines the trustworthiness and responsibility of these models. This talk will explore methods to navigate the misalignment problem by addressing three key research questions: (1) How to mitigate the risk of a misaligned LLM when only limited inference access is available? (2) How to ensure a reliable alignment process in multimodal scenarios? (3) How to integrate missing or customized alignment objectives to achieve precise control over model behavior in diverse contexts.

..................................................................................11/21/2024..................................................................................

Title: Style Outweighs Substance: Failure Modes of LLM Judges in Alignment Benchmarking

Abstract: The release of ChatGPT in November 2022 sparked an explosion of interest in LLM alignment with human values, preferences and standards. Existing methods claim superiority by virtue of better correspondence with human pairwise preferences, often measured by LLM judges. But do LLM-judge preferences translate to progress on other, more concrete metrics for alignment, and if not, why not? Recent joint research with NYU, Columbia and Arthur.AI shows that (1) LLM-judgments do not correlate with concrete measures of safety, world knowledge, and instruction following; (2) LLM judges have powerful implicit biases, prioritizing style over factuality and safety; and (3) the supervised fine-tuning (SFT) stage of post-training, rather than RLHF, has the greatest impact on objective measures of alignment.

..................................................................................9/24/2024..................................................................................

Title: Connections Between Approximate Inference and Adaptive Optimization

Abstract: He will talk about designing training methods for large neural networks.

..................................................................................4/26/2024..................................................................................

Title: AIScientist: Toward Automated Literature Understanding and Scientific Discovery

Abstract: Due to the rapid growth of publications varying in quality, there exists a pressing need to help scientists digest and evaluate relevant papers, thereby facilitating scientific discovery. This talk aims to provide an overview of the AI-assisted scientific paper lifecycle, detailing how machines can augment every stage of the research process for the scientist, including scientific literature understanding, experiment development, manuscript draft writing, and finally draft evaluation. I will first cover fine-grained few-shot scientific entity extraction and then show how the extracted knowledge graph can be applied to scientific literature review and scientific hypothesis discovery.