Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Learn about HbBoPs, a novel framework that combines structural-aware Gaussian process surrogates with Hyperband scheduling to efficiently select optimal prompts for large language models in black-box settings. Discover how this approach addresses the combinatorial optimization challenge of automated prompt selection when models are only accessible via APIs, where evaluation costs are high and gradients are unavailable. Explore the framework's key innovation of representing prompts through embeddings of instructions and exemplars as modular components, enabling the surrogate model to capture structural similarities across different prompts. Understand how Hyperband's multi-fidelity scheduling adaptively allocates computational budget by evaluating prompts at varying fidelity levels, significantly reducing the number of expensive full prompt evaluations required. Examine validation results across ten datasets and three different large language models, demonstrating consistent performance improvements over state-of-the-art baselines and competing methods. Gain insights into how this work extends the scope of AutoML by applying multi-fidelity Bayesian optimization techniques, traditionally used for hyperparameter tuning and neural architecture search, to the structured, high-cost decision spaces characteristic of modern natural language processing applications.
