Adaptive Acquisition Selection for Bayesian Optimization with Large Language Models
Giang Ngo, Dat Phan Trong, Dang Nguyen, Sunil Gupta, Svetha Venkatesh
TL;DR
This paper introduces LMABO, a novel framework that uses a pre-trained Large Language Model to adaptively select acquisition functions in Bayesian Optimization, significantly improving performance across diverse problems.
Contribution
The paper proposes LMABO, the first framework leveraging LLMs as zero-shot strategists for adaptive acquisition selection in Bayesian Optimization.
Findings
LMABO outperforms static and other adaptive methods on 50 benchmark problems.
The LLM effectively synthesizes optimization state to adapt acquisition strategies.
LMABO demonstrates the advantage of using rich, real-time information in BO.
Abstract
Bayesian Optimization critically depends on the choice of acquisition function, but no single strategy is universally optimal; the best choice is non-stationary and problem-dependent. Existing adaptive portfolio methods often base their decisions on past function values while ignoring richer information like remaining budget or surrogate model characteristics. To address this, we introduce LMABO, a novel framework that casts a pre-trained Large Language Model (LLM) as a zero-shot, online strategist for the BO process. At each iteration, LMABO uses a structured state representation to prompt the LLM to select the most suitable acquisition function from a diverse portfolio. In an evaluation across 50 benchmark problems, LMABO demonstrates a significant performance improvement over strong static, adaptive portfolio, and other LLM-based baselines. We show that the LLM's behavior is a…
Peer Reviews
Decision·ICLR 2026 Poster
The work tackles a meta design choice in BO/AutoML—which AF to use—rather than only tuning model hyperparameters. In practice this could translate in a more robust “out‑of‑the‑box” optimizer that requires less trial to find the optima. The paper reports extensive experiments indicating that the meta decision is relevant. It is important to have flexibility in what AF to choose along the acquisition process.
Presentation. I think the presentation and how the paper is written is not good in general. To give some examples. 1) Some figures are hard to read or to use for significance judgments. In Figure 2, the variance of the result is wide, making it difficult to visually assess the significance of the results. 2) Table 2 uses cryptic labels (“LMABO‑AB1…AB4”); these should be replaced or augmented with descriptive names. Novelty. The idea of leveraging LLM “knowledge” to steer BO is not brand‑new (e.
**Strong results.** Casting AF selection itself as an in‑context decision problem for an LLM shows that the meta decision of selecting the acquisition function is highly relevant for the whole acquisition process. The paper shows strong results.
**Novelty.** While the framing is neat, the algorithmic contribution largely reduces to a prompt + a state serialization + a portfolio choice. Many recent works have leveraged LLM inductive biases for decision selection or candidate generation in BO, so the conceptual step—“use an LLM to pick an AF given a textual state”—feels incremental without additional design elements. **What paper should have.** The paper positions an LLM as the decision-maker for selecting the acquisition function, with
* The paper introduces a framework that successfully recasts the task of acquisition function (AF) selection in Bayesian Optimization (BO) as an in-context decision-making problem. This approach leverages a pre-trained Large Language Model (LLM) to select the most appropriate AF at each optimization step based on its implicitly encoded knowledge of optimization principles. * A primary technical contribution is the design of a structured state representation that translates the complex, multi-f
* The LMABO algorithm requires calling the Large Language Model at every single optimization step. Although the financial cost per token is very low, the overall economic feasibility for extensive or complex problems using commercial LLM APIs requires more detailed analysis. * The framework's high performance is fundamentally dependent on the sophisticated reasoning capabilities of the underlying LLM. The core implementation relies predominantly on Gemini-2.5 Flash. Ablation studies confirm this
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Bandit Algorithms Research · Machine Learning and Data Classification · Gaussian Processes and Bayesian Inference