The Seventh FEVER Workshop

Language Models (LMs), trained as generalist systems have made it much easier to prototype impressive AI demos, but turning LMs into reliable AI systems remains challenging, as their monolithic nature makes them hard to control, debug, and improve. To tackle this, the AI community is increasingly building Compound AI Systems, i.e. modular programs that uses LMs as specialized components, but most such systems are highly brittle in practice: they couple task decomposition with choices about prompting, finetuning, inference-time strategies, and even individual LMs. Instead of prompting LMs with string templates, what if we could build more reliable and scalable AI systems by programmatically composing natural-language-typed modules that can learn from data? This is the key idea behind the DSPy framework, which introduces natural-language programming abstractions (DSPy Signatures and Modules) and a set of new ML algorithms (DSPy Optimizers) that compile high-level programs down into optimized prompting and finetuning strategies for Compound AI Systems. In this talk, we will provide an overview of what DSPy is and how the DSPy programming model works, as well as a deeper dive into the latest automatic prompt optimization algorithm that powers DSPy, MIPROv2.

False information is shared everyday and everywhere and in countless forms. Traditionally, such information was produced and disseminated by people, with various intentions ranging from benign exaggerations to deliberate manipulation. Over the past two decades, research in AI has shown that algorithms can often outperform humans in identifying falsehood, providing a valuable tool in areas such as security, media, and consumer protection. However, the recent rise of large language models adds complexity to this landscape: these models have now become remarkably proficient at generating false information, often convincingly so. The challenge is further amplified by the large number of languages and cultural nuances worldwide, making it increasingly difficult to detect and manage false content. So, where do we go from here?

Fact-checkers, researchers and policymakers face challenges accessing accurate information online on key topics in many African contexts. To overcome these challenges, fact-checkers develop short-term tactics and see a need for longer-term strategies. Using a model developed from examination of fact-checks thus produced, I argue it is possible to better focus AI and research efforts on the potential of specific false claims to cause, or not to cause, or contribute to specific substantive negative consequences, or harms, for individuals and society, and discuss the implications.

In this talk, I will provide an overview of trends we've observed in online misinformation by analyzing years of online fact-checking data from the International Fact-Checking Network's (IFCN) ClaimReview datasets. The focus will be on recent trends in visual misinformation, covering new genAI based misinformation, "cheap fakes," and misleading context manipulations. I'll also share some surprising statistics that challenge common beliefs about the most prevalent types of misinformation. Beyond identifying trends, I'll discuss mitigation strategies, including how we can improve information literacy tools, the opportunities and limitations of using AI to detect manipulated content, how various provenance methods together with AI can help mitigate out-of-context manipulations, and new opportunities for multimodal LLMs in the space of visual fact-checking.