Computer Science & Engineering Building (CSE), Room 1242

The applications of large language models (LLMs) are increasingly complex and diverse, necessitating efficient and reliable frameworks for building and deploying them. In this talk, I will begin with algorithms and systems for serving LLMs for everyone (FlexGen, S-LoRA, VTC), highlighting the growing trend of personalized LLM services. My work addresses the need to run LLMs locally for isolated individual needs. It also tackles the problem of efficiency and service fairness when resource sharing among many users is required. Once we have efficient deployment, a primary concern is the reliability of generation. The second part of this talk aims to address this issue by exploring verifiable code generation. To achieve this, I adopt tools in formal verification to facilitate LLMs in generating correctness certificates alongside other artifacts (Clover). Finally, I will touch on future research avenues, such as integrating formal methods with LLMs and developing programming systems for generative AI.

Abstract: Reasoning, the ability to logically draw conclusions from existing knowledge, has been long pursued as a goal of artificial intelligence. Although numerous learning algorithms have been developed for reasoning, most of them are limited to the domain they are trained on. By contrast, humans often derive high-level rules or principles from experience and apply them to new domains — an ability referred as inductive generalization. In this talk, we present a series of works that learn inductive representations for reasoning over knowledge graphs. First, we introduce Neural Bellman-Ford Networks (NBFNet) that captures paths between entities and can generalize to graphs of new entities. Then we discuss Graph Neural Network Query Executor (GNN-QE), an extension of NBNet that answers multi-hop logical queries and generalizes well on our inductive benchmark. Finally, by learning inductive representations for both entities and relations, we demonstrate that a model can generalize to any graph with arbitrary entity and relation vocabularies, paving the way for foundation models for knowledge graph reasoning.