In-Context Black-Box Optimization with Unreliable Feedback
Researchers have introduced a novel approach called Feedback-Informed In-Context Black-Box Optimization (FICBO) to address challenges in scientific and engineering optimization. Traditional black-box optimization often utilizes side information from experts or simulators, which can accelerate search but may also be biased or misleading. Existing methods typically handle single tasks or ignore auxiliary feedback at test time. FICBO overcomes these limitations by employing a pretrained optimizer that conditions on both observed history and cheap auxiliary feedback for current candidate sets. The team developed a structured feedback prior to model how different feedback sources vary in access, relevance, and distortion relative to the true objective. This prior is used to pretrain a feedback-aware transformer. At test time, the model dynamically estimates source reliability by comparing observed objective values with auxiliary signals, thereby improving query selection. Empirical results on synthetic and real-world tasks demonstrate that FICBO effectively exploits informative feedback while maintaining robustness against weak or misleading sources, outperforming existing baselines. The study also provides insights into the model's interpretability and decision-making processes regarding test-time sources.
Wire timeline
In-Context Black-Box Optimization with Unreliable Feedback
Researchers have introduced a novel approach called Feedback-Informed In-Context Black-Box Optimization (FICBO) to address challenges in scientific and engineering optimization. Traditional black-box optimization often utilizes side information from experts or simulators, which can accelerate search but may also be biased or misleading. Existing methods typically handle single tasks or ignore auxiliary feedback at test time. FICBO overcomes these limitations by employing a pretrained optimizer that conditions on both observed history and cheap auxiliary feedback for current candidate sets. The team developed a structured feedback prior to model how different feedback sources vary in access, relevance, and distortion relative to the true objective. This prior is used to pretrain a feedback-aware transformer. At test time, the model dynamically estimates source reliability by comparing observed objective values with auxiliary signals, thereby improving query selection. Empirical results on synthetic and real-world tasks demonstrate that FICBO effectively exploits informative feedback while maintaining robustness against weak or misleading sources, outperforming existing baselines. The study also provides insights into the model's interpretability and decision-making processes regarding test-time sources.
cs.AI updates on arXiv.org