LILO: Bayesian Optimization with Natural Language Feedback
Researchers have introduced Language-in-the-Loop Optimization (LILO), a novel Bayesian optimization framework designed to address complex, subjective real-world problems that lack explicit closed-form objectives. LILO utilizes a large language model (LLM) to translate free-form natural language feedback and prior knowledge from decision-makers into structured preference signals, surpassing the limitations of traditional scalar or pairwise feedback methods. By integrating these LLM-derived preferences into a Gaussian process proxy model, the framework enables principled, acquisition-driven exploration with calibrated uncertainty. Crucially, LILO positions the LLM in a supporting role rather than as the primary optimizer, thereby preserving the sample efficiency and stability inherent to Bayesian optimization while offering a flexible user interface. Benchmark tests on both synthetic and real-world datasets demonstrate that LILO consistently outperforms conventional preference-based Bayesian optimization methods and LLM-only optimizers. The approach shows particularly significant improvements in regimes where feedback is limited. This research was presented at the 43rd International Conference on Machine Learning in Seoul, South Korea, marking a significant advancement in combining natural language processing with robust optimization techniques.
Wire timeline
LILO: Bayesian Optimization with Natural Language Feedback
Researchers have introduced Language-in-the-Loop Optimization (LILO), a novel Bayesian optimization framework designed to address complex, subjective real-world problems that lack explicit closed-form objectives. LILO utilizes a large language model (LLM) to translate free-form natural language feedback and prior knowledge from decision-makers into structured preference signals, surpassing the limitations of traditional scalar or pairwise feedback methods. By integrating these LLM-derived preferences into a Gaussian process proxy model, the framework enables principled, acquisition-driven exploration with calibrated uncertainty. Crucially, LILO positions the LLM in a supporting role rather than as the primary optimizer, thereby preserving the sample efficiency and stability inherent to Bayesian optimization while offering a flexible user interface. Benchmark tests on both synthetic and real-world datasets demonstrate that LILO consistently outperforms conventional preference-based Bayesian optimization methods and LLM-only optimizers. The approach shows particularly significant improvements in regimes where feedback is limited. This research was presented at the 43rd International Conference on Machine Learning in Seoul, South Korea, marking a significant advancement in combining natural language processing with robust optimization techniques.
cs.AI updates on arXiv.org