DASH: Decoupled Adaptive Surrogate-Acquisition Harness for Automated Bayesian Optimization
Original title:DASH: Decoupled Adaptive Surrogate - Acquisition Harness for Automated Bayesian Optimization
AI Summary
DASH is an LLM-enhanced framework for automated Bayesian optimization that decouples surrogate selection from acquisition adaptation. Surrogates are chosen using predictive reliability, uncertainty calibration, and ranking consistency, while a two-stage controller reallocates quotas among acquisition functions and lets an LLM make the final choice from a BO shortlist. The framework also adds knowledge-guided warm starts and structured memory. Across four chemical optimization tasks, the paper reports a 12.51% improvement in trajectory-level Acceleration Factor and a 5.00% improvement in endpoint Enhancement Factor over the strongest AutoBO baseline. Ablations and contamination checks are also reported.
Why it's worth reading
The paper reframes AutoBO component selection as two distinct control problems and reports quantified gains on chemical optimization tasks, together with ablations and benchmark-contamination checks.
Deep Read
What happened
Original facts: The paper introduces DASH, a Decoupled Adaptive Surrogate-Acquisition Harness for LLM-enhanced automated Bayesian optimization. It separates surrogate-model selection from acquisition-function adaptation and adds knowledge-guided warm starts plus structured memory.
Core technology
Original facts: DASH selects surrogates using predictive reliability, uncertainty calibration, and ranking consistency. Its acquisition controller periodically reallocates quotas across acquisition functions, constructs a BO shortlist, and delegates the final candidate choice to an LLM.
Key evidence and numbers
Original facts: On four chemical optimization tasks, the paper reports a 12.51% improvement in trajectory-level Acceleration Factor and a 5.00% improvement in endpoint Enhancement Factor over the best AutoBO baseline. The abstract states that results remain strong across LLM backbones. It also reports ablations supporting complementary contributions from the components, plus full-table and behavioral contamination checks with no detectable evidence that direct benchmark memorization or source-cell leakage explains the gains.
Why it matters
Analysis: Surrogate selection asks whether predictions are trustworthy, while acquisition adaptation asks how the campaign should balance exploration and exploitation at its current stage. These objectives differ. DASH therefore provides a more interpretable control decomposition and may make failures easier to attribute to modeling reliability or decision policy.
Practical impact
Analysis: In costly chemical optimization with sparse feedback and useful domain knowledge, warm starts and structured memory could reduce cold-start waste. For implementation, the design also suggests logging surrogate diagnostics, acquisition-quota changes, LLM shortlist decisions, and experimental feedback separately instead of retaining only the final trajectory.
Limitations and uncertainty
Original facts: The abstract identifies four chemical optimization tasks but does not specify task names, dataset sizes, baseline configurations, variance, statistical significance, runtime, or LLM-call budget. Analysis: The available evidence is insufficient to establish transfer to non-chemical domains, real wet-lab campaigns, or higher-dimensional search. Unverified inference: Final LLM choices may be sensitive to prompt templates, candidate ordering, and memory formatting; this requires confirmation from the full paper's robustness experiments.
Original sources
- arXiv abstract page
- Paper: DASH: Decoupled Adaptive Surrogate - Acquisition Harness for Automated Bayesian Optimization
- arXiv ID: 2608.00641
- Published: 2026-08-01T12:53:34.000Z