文章背景与核心概要
DASH (Decoupled Adaptive Surrogate-Acquisition Harness) 是一种旨在提升自动化贝叶斯优化 (AutoBO) 性能的创新框架。传统的 AutoBO 方法往往受限于代理模型与采集函数之间的不匹配问题，而 DASH 通过将这两个核心组件解耦，实现了更灵活的优化策略。
该框架的核心在于三方面：基于预测可靠性和校准性的代理模型选择、利用两阶段控制器及大语言模型 (LLM) 进行动态采集函数分配，以及通过知识引导的冷启动和结构化记忆来整合领域经验。在化学优化任务的实验中，DASH 展现了显著的性能提升，证明了其在复杂搜索空间中的高效性与鲁棒性。
DASH：用于自动化贝叶斯优化的解耦自适应代理-采集框架
作者： Changquan Zhao, Yuxiang Sun, Ruihao Zhu, Cheng Hua, Yulian He
arXiv： 2608.00641 [cs.AI]
提交日期： 2026年8月1日 (v1), 2026年8月6日 (v2)
摘要
DASH (Decoupled Adaptive Surrogate-Acquisition Harness) is a novel framework designed to enhance Automated Bayesian Optimization (AutoBO). Traditional AutoBO methods often struggle with the mismatch between surrogate models and acquisition functions. DASH addresses this by decoupling these components: * Surrogate Selection: Based on predictive reliability, uncertainty calibration, and ranking consistency. * Acquisition Control: A two-stage controller that reallocates quotas across functions and utilizes an LLM for final selection. * Integrated Harness: Incorporates knowledge-guided warm starts and structured memory to ground optimization in domain-specific feedback.
Experimental results across four chemical optimization tasks demonstrate that DASH outperforms existing baselines by 12.51% in trajectory-level Acceleration Factor and 5.00% in endpoint Enhancement Factor.
核心功能
1. 解耦架构
与以往强制联合选择代理-采集对的方法不同，DASH 将它们视为独立的实体。这使得系统能够在优化代理模型预测精度的同时，独立地根据优化过程的不同阶段调整采集策略。
Unlike previous methods that force a joint selection of surrogate-acquisition pairs, DASH treats them as distinct entities. This allows the system to optimize the surrogate for predictive accuracy while independently adapting the acquisition strategy to the specific stage of the optimization campaign.
2. LLM 增强决策
DASH 利用大语言模型 (LLM) 来管理采集函数的最终选择，确保优化过程保持灵活性和上下文感知能力。
DASH leverages Large Language Models (LLMs) to manage the final selection of acquisition functions, ensuring that the optimization process remains flexible and context-aware.
3. 领域基础优化
通过利用“知识引导的冷启动”和“结构化记忆”，该框架有效地整合了历史数据和领域专业知识，从而在复杂的搜索空间中实现了更快的收敛和更稳健的性能。
By utilizing a "knowledge-guided warm start" and "structured memory," the framework effectively incorporates historical data and domain expertise, leading to faster convergence and more robust performance in complex search spaces.
性能指标
- 加速因子 (Acceleration Factor)： 较最佳 AutoBO 基准提升 12.51%。
- 增强因子 (Enhancement Factor)： 终点结果提升 5.00%。
- 鲁棒性： 在不同的 LLM 主干模型下性能保持一致，且无基准测试记忆或数据泄露的迹象。
- Acceleration Factor: 12.51% improvement over the best AutoBO baseline.
- Enhancement Factor: 5.00% improvement in endpoint results.
- Robustness: Performance remains consistent across different LLM backbones, with no evidence of benchmark memorization or data leakage.
访问与资源
