Large Language Models (LLMs) frequently suffer from knowledge-intensive questions, often being inconsistent by providing different outputs despite given the same input. The response quality worsens when the user expresses a firm opposing stance which causes the LLMs to adjust its response despite the correct initial one. These behaviors decrease the reliability and validity of the responses provided by these models. In this paper, we attempt to 1) raise awareness of the inherent risks that follow from overly relying on AI agents like ChatGPT by showing how Chain-of-Feedback (CoF) triggers LLMs to deviate more from the actual answer and 2) suggest a novel prompting method, Recursive Chain of Feedback (R-CoF), that we are conducting further study. The CoF system takes in an open-ended multi-step question. Then, we repetitively provide meaningless feedback requesting another attempt. Our preliminary experiments show that such feedback only decreases the quality of the response. On the other hand, to mitigate the effects of the aforementioned inconsistencies, we present a novel method of recursively revising the initial incorrect reasoning provided by the LLM by repetitively breaking down each incorrect step into smaller individual problems.
Mining large corpora can generate useful discoveries but is time-consuming for humans. We formulate a new task, D5, that automatically discovers differences between two large corpora in a goal-driven way. The task input is a problem comprising a research goal "$\textit{comparing the side effects of drug A and drug B}$" and a corpus pair (two large collections of patients' self-reported reactions after taking each drug). The output is a language description (discovery) of how these corpora differ (patients taking drug A "$\textit{mention feelings of paranoia}$" more often). We build a D5 system, and to quantitatively measure its performance, we 1) contribute a meta-dataset, OpenD5, aggregating 675 open-ended problems ranging across business, social sciences, humanities, machine learning, and health, and 2) propose a set of unified evaluation metrics: validity, relevance, novelty, and significance. With the dataset and the unified metrics, we confirm that language models can use the goals to propose more relevant, novel, and significant candidate discoveries. Finally, our system produces discoveries previously unknown to the authors on a wide range of applications in OpenD5, including temporal and demographic differences in discussion topics, political stances and stereotypes in speech, insights in commercial reviews, and error patterns in NLP models.