Visit
Thu 05/02
Ruiqi Zhong headshot

Visitor Talk: Explaining Datasets with Language Models

How do online discussions think about policy? What kinds of images are more likely to cause model errors? What types of images are more memorable for humans? Answering these types of questions is valuable for engineering, science, and policy purposes, but they are also challenging for humans because it requires discovering and explaining patterns from large datasets. Ruiqi Zhong proposes to explain datasets with LLMs using natural language, an efficient interface for humans to understand complex patterns. To account for more complex dataset structures, Zhong formalizes a new family of statistical models parameterized with natural language strings; additionally, Zhong develops a general optimization framework based on continuous relaxation to learn these models efficiently. At the end of this talk, Zhong will go through a few examples of real-world applications, including explaining model internal representations (used by OpenAI & Anthropic), finding failures of multi-modal models (Midjourney 5.1, DALL-E, VideoFusion), and explaining human preference on ChatBotArena (that ranks SOTA systems such as GPT-4, Gemini, Claude, LLama-3).

Speaker Bio

Ruiqi Zhong is a Ph.D. student at UC Berkeley, co-advised by Jacob Steinhardt and Dan Klein. He works on NLP and focuses on helping humans understand complex datasets and model behaviors. He is also a part-time researcher at Anthropic, working on AI Safety. Outside of research, he is interested in the philosophy of science, international relations, and baking.