Feedback-Informed In-Context Black-Box Optimization Method

A novel technique known as feedback-informed in-context black-box optimization (FICBO) has been unveiled in a preprint on arXiv (2605.06187v1). This method tackles black-box optimization challenges when side information from experts, simulators, or heuristics is present but potentially unreliable. Conventional feedback-aware Bayesian optimization approaches focus on one task at a time, which restricts the ability to generalize across various feedback sources. While in-context optimizers can adapt across tasks, they often depend solely on optimization history during testing. FICBO utilizes both observed history and inexpensive auxiliary feedback for the current candidate set. Researchers created a structured feedback prior to model variations in access, relevance, and distortion of feedback sources relative to the true objective, aiming to enhance search efficiency in science and engineering fields where feedback may be biased or misleading.