演講資訊

The Future of Personalized Universal Assistant
09:10-10:00

We've moved way beyond the old days of building discovery, recommendation, decision support, and other AI tools using traditional ML and pattern recognition techniques. The future of universal personal assistance for discovery and learning is upon us. How will multimodality image, video, and audio understanding, and reasoning abilities of large foundation models change how we build these systems? I will shed some initial light on this topic by discussing 3 trends: First, the move to a single multimodal large model with reasoning abilities; Second, the fundamental research on personalization and user alignment; Third, the combination of System 1 and System 2 cognitive abilities into a single universal assistant. 

Factual and Faithful Generation with Retrieval
15:10-15:50
Despite demonstrating remarkable flexibility and capability, large language models (LLMs) frequently produce responses that are inconsistent or factually inaccurate. Recently, retrieval-augmented generation (RAG) has emerged as a popular mitigating strategy by equipping LLMs with relevant knowledge, but it still falls short of providing a comprehensive solution. In this presentation, I will discuss our recent efforts to enhance the factual accuracy and faithfulness of LLMs. The talk will begin with an overview of retrieval-augmented methods and automatic evaluation metrics for assessing the factuality of LLM responses, followed by a factuality-aware alignment method, comprising factuality-aware SFT and factuality-aware RL through direct preference optimization. Next, I will discuss how online retrieval and fact-checking feedback can be leveraged to improve the factuality of model responses, and also how fine-grained, sentence-level citations can be generated.  Finally, I will conclude by discussing key open problems and outlining directions for addressing them in the near future.

From Distillation to Personalization: Accelerating and Customizing Generative AI Models
16:00-16:40

The convergence of language, vision, and generative models is a captivating and rapidly advancing domain for AI applications. In this talk, we explore the evolving landscape of generative AI through the lenses of model distillation and personalization. We begin with recent breakthroughs in distillation techniques that dramatically reduce the size and inference cost of large generative models while preserving their capabilities. Building on this, we delve into emerging methods for model personalization, enabling fine-grained customization of generative behavior to suit individual users, tasks, or domains. By connecting these two threads, we highlight how the combination of efficiency and adaptability is shaping the next generation of practical, scalable, and user-centric generative AI systems. 

Toward Efficient Planning in Reinforcement Learning with Learned Options
10:20-11:40

Planning with options -- a sequence of primitive actions -- has been shown effective in reinforcement learning within complex environments. However, most prior work have focused on planning with predefined options or learned options through expert demonstration data. In this talk, I will first review the key challenges of integrating Option into the RL planning process. Then, I will introduce our recent work, OptionZero, which incorporates an option network into MuZero, providing autonomous discovery of options through self-play games. Our findings shows that the agent not only learns options but also acquires strategic skills, showing promising directions for option discovery and utilization in future planning.

Trapdoor-based Defense against Privacy Attacks
10:20-11:40

Model Inversion (MI) attacks threaten data privacy by reconstructing training data from AI models. In this talk, I will present Trap-MID, a defense that embeds a trapdoor trigger into the model, causing MI attacks to recover the trapdoor pattern instead of private data. Unlike traditional regularization-based defenses, Trap-MID uses deception to mislead attackers. I will discuss theoretical insights on trapdoor effectiveness and naturalness, along with empirical results showing state-of-the-art performance against various MI attacks—achieved without extra data or significant overhead.

Opportunities, Challenges, and Future Directions of Large Language Models in Mathematics Education
10:20-11:40

Recent advances in large language models (LLMs) have opened new possibilities for their application in mathematics education. This talk explores the potential of LLMs to support students' problem-solving processes, diagnose errors, and deliver personalized feedback—while also critically examining current limitations and future research directions. We will discuss (1) the capabilities and shortcomings of LLMs in understanding and evaluating students' reasoning, (2) their effectiveness in generating adaptive and pedagogically meaningful feedback, and (3) their role in identifying ambiguity and errors in math word problems, including self-optimization mechanisms. Through these perspectives, the talk aims to assess the practical utility and developmental prospects of LLMs as intelligent educational assistants. 

Algorithms and Applications of Programmatic Reinforcement Learning (NeurIPS 2024 Tutorial)
12:30-13:20

In this tutorial, we will discuss recent advances in program synthesis that enable the generation of programmatic policies for reinforcement learning (RL). Instead of representing policies using deep neural networks, programmatic RL (PRL) methods aim to synthesize program policies structured in a human-readable domain-specific language. PRL reformulates the RL into learning to write a program that can be executed in an environment and maximize the return, potentially yielding improved interpretability and generalizability. We will cover different families of algorithms that rely on search and learning-based methods, including those using large language models to help with the search for programmatic policies. This tutorial was selected to be presented at NeurIPS 2024 in Vancouver, Canada.

Beyond imitation, exploration for robot learning
13:30-14:50

Imitation learning has made great progress in learning robot policies based on a simple idea: training the robot policy to mimic human's behaviors from datasets of expert demonstrations.  However,  curation of large-scale datasets is costly, requiring skilled teleoperators and substantial operation time.  Worse, the imitation learning paradigm overlooks interactions without the environment, creating a huge knowledge gap between the environmental dynamics and the embodied robots.  In this talk, I will discuss exploration--the key to address challenges in data curation and decision making in complex and dynamic environments.  I will discuss how exploration could facilitate task planning in Large Language Models, meanwhile, how exploration could enable zero-shot data generation for imitation learning.  Moreover, I will present how exploration could enable sample efficient reinforcement learning.

Toward Human-Centered Physical AI
13:30-14:50

Recent years have seen remarkable advancements in generative AI and its applications across various domains. However, these successes have largely remained in the digital realm and are yet to fully extend into the physical world. There is growing interest in the research community to explore how modern AI techniques can empower robotic systems. In this talk, I will share my experience in developing AI-powered physical systems—what we refer to as Physical AI. I will also introduce our recent efforts aimed at enabling these systems to augment human capabilities, moving us closer to truly human-centered physical intelligence.

Toward Personalized Precision Medicine in Antibiotic Resistance: Advancing from Predictive and Generative to Agentic AI
13:30-14:50

Antibiotic resistance is a pressing global health issue, leading to higher patient mortality, longer hospital stays, and increased healthcare costs. To tackle this, we need rapid diagnostic tools and innovative therapies tailored to individual patients. Artificial intelligence (AI) has the potential to transform this field by enabling personalized precision medicine. Predictive AI methods combined with matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) have improved pathogen identification and antibiotic susceptibility predictions, allowing for timely treatment. Beyond predictions, generative AI has aided in discovering and optimizing novel antimicrobial peptides (AMPs) that could combat resistant pathogens. Although these AI-generated peptides show promise, further experimental validation is necessary to understand their effectiveness and resistance profiles. Looking ahead, we envision agentic AI systems that can provide real-time, adaptive therapeutic interventions adapted to patient-pathogen interactions. This highlights AI's evolving role in developing a comprehensive framework to address antibiotic resistance.