In this work we study prompt optimization for Language Model Programs (ie. pipelines composed of one or more modular language model calls which are often chained together). We introduce several strategies to solve the challenges of creating high quality prompts and performing credit assignment across multiple modules in this setting. Using these strategies, we develop MIPRO, a novel LM Program optimizer which outperforms baselines on five of six diverse LM programs. We open source both MIPRO and our benchmark on DSPy.

In this work we explore the potential of multimodal foundation models for automating enterprise workflows. Our proposed system, ECLAIR, leverages these models to understand and execute workflows with high accuracy and minimal setup. Our initial experiements demonstrate that multimodal FMs can address the limitations of traditional RPA, providing a promising direction for the future of enterprise workflow automation.

This study focuses on the challenge of maintaining model fairness under distribution shifts in healthcare machine learning applications. We adopt a causal framing to motivate conditional independence tests as a key tool for characterizing distribution shifts. Through two medical case studies, we demonstrate how understanding the structure of these shifts can diagnose and mitigate fairness failures, often more complex than previously recognized. We propose potential remedies across various stages of the machine learning pipeline to maintain fairness in real-world settings.