Evidence-Gated LLM Priors for Multi-Objective Bayesian Optimization

The paper presents a novel approach for integrating large language model (LLM) priors into multi-objective Bayesian optimization through an objective-wise reputation-market mechanism. This method dynamically updates expert weights based on observed feedback and introduces a decoupled counterfactual gate, allowing for varying levels of reliance on LLM priors. The findings indicate that dynamic calibration enhances robustness, but LLM confidence is not consistently beneficial across different benchmarks, highlighting the need for careful consideration of expert contributions in optimization tasks.