46 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Evaluating factuality in long-form generation
University of Maryland, College Park
Recording
Abstract
Evaluating the factuality of long-form texts generated by large language models, such as articles, essays and books, is a growing problem: as long-form writing tasks become longer and more complex, how can progress even be measured? The talk first proposes a high-level framework, applicable to both human and automatic evaluation, that decomposes a long-form text into simpler atomic units and then evaluates each unit on a specific aspect, and demonstrates its effectiveness for evaluating factuality and faithfulness in tasks such as long-form question answering and book summarization, along with ways to speed up the pipeline to scale evaluation. It then turns to evaluation of agentic writing pipelines, in which models use tools and write and execute code to help produce long-form reports, focusing on the AnalystBench project. It concludes with open problems: reliance on faulty retrievers and search engines, verifying complex, highly compositional claims, and evaluating interpretive rather than fully objective claims.