Generated texts from large language models (LLMs) have been shown to exhibit a variety of harmful, human-like biases against various demographics. These findings motivate research efforts aiming to understand and measure such effects. This paper introduces a causal formulation for bias measurement in generative language models. Based on this theoretical foundation, we outline a list of desiderata for designing robust bias benchmarks. We then propose a benchmark called OccuGender, with a bias-measuring procedure to investigate occupational gender bias. We test several state-of-the-art open-source LLMs on OccuGender, including Llama, Mistral, and their instruction-tuned versions. The results show that these models exhibit substantial occupational gender bias. Lastly, we discuss prompting strategies for bias mitigation and an extension of our causal formulation to illustrate the generalizability of our framework.

Users can divulge sensitive information to proprietary LLM providers, raising significant privacy concerns. While open-source models, hosted locally on the user’s machine, alleviate some concerns, models that users can host locally are often less capable than proprietary frontier models. Toward preserving user privacy while retaining the best quality, we propose Privacy-Conscious Delegation, a novel task for chaining API-based and local models. We utilize recent public collections of user-LLM interactions to construct a natural benchmark called PUPA, which contains personally identifiable information (PII). To study potential approaches, we devise PAPILLON, a multi-stage LLM pipeline that uses prompt optimization to address a simpler version of our task. Our best pipeline maintains high response quality for 85.5% of user queries while restricting privacy leakage to only 7.5%. We still leave a large margin to the generation quality of proprietary LLMs for future work.

Many low-resource languages, such as Prakrit, present significant linguistic complexities and have limited modern-day resources. These languages often have multiple derivatives; for example, Prakrit, a language in use by masses around 2500 years ago for 500 years, includes Pali and Gandhari, which encompass a vast body of Buddhist literature, as well as Ardhamagadhi, rich in Jain literature. Despite these challenges, these languages are invaluable for their historical, religious, and cultural insights needed by non-language experts and others.To explore and understand the deep knowledge within these ancient texts for non-language experts, we propose a novel approach: translating multiple dialects of the parent language into a contemporary language and then enabling them to interact with the system in their native language, including English, Hindi, French and German, through a question-and-answer interface built on Large Language Models. We demonstrate the effectiveness of this novel AI-Tutor system by focusing on Ardhamagadhi and Pali.