💻 computer science

LABO: LLM-Accelerated Bayesian Optimization through Broad Exploration and Selective Experimentation

The paper introduces LABO, a novel Bayesian optimization framework that enhances scientific discovery efficiency by dynamically balancing low-cost LLM predictions for broad exploration with selective, costly real-world experiments for high-uncertainty regions, thereby achieving superior sample efficiency and theoretical regret bounds compared to existing methods.

Original authors: Zhuo Chen (equal contribution), Xinzhe Yuan (equal contribution), Jianshu Zhang (Shanghai Artificial Intelligence Laboratory, Shanghai, China, School of Computer Science, Shanghai Jiao Tong University

Published 2026-05-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhuo Chen (equal contribution), Xinzhe Yuan (equal contribution), Jianshu Zhang (Shanghai Artificial Intelligence Laboratory, Shanghai, China, School of Computer Science, Shanghai Jiao Tong University, Shanghai, China), Jinzong Dong (Shanghai Artificial Intelligence Laboratory, Shanghai, China, School of Automation, Central South University, Changsha, China), Ruichen Zhou (College of New Energy and Materials, China University of Petroleum, Beijing, China), Yingchun Niu (College of New Energy and Materials, China University of Petroleum, Beijing, China), Tianhang Zhou (College of Carbon Neutrality Future Technology, China University of Petroleum, Beijing, China), Yu Yang Fredrik Liu (DeepVerse PTE. LTD., Singapore), Yuqiang Li (Shanghai Artificial Intelligence Laboratory, Shanghai, China), Nanyang Ye (Shanghai Artificial Intelligence Laboratory, Shanghai, China, School of Computer Science, Shanghai Jiao Tong University, Shanghai, China), Qinying Gu (Shanghai Artificial Intelligence Laboratory, Shanghai, China)

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a chef trying to invent the world's perfect sandwich. You have a limited budget for ingredients, and every time you make a sandwich to taste-test it, it costs you $100 and takes all day. You can't just make thousands of sandwiches to find the best one; you have to be smart about which ones you actually cook.

This is the problem scientists face when designing new drugs, materials, or batteries. The "experiments" are expensive and slow. This is where LABO comes in.

Here is how the paper explains LABO, using simple analogies:

1. The Problem: The "Cold Start" and the "Empty Wallet"

Traditional methods (called Bayesian Optimization) are like a chef who only tastes the sandwich after baking it.

The Cold Start: At the very beginning, the chef has no idea what ingredients work well. They have to guess blindly, wasting expensive ingredients.
The Budget: Since every taste test costs $100, the chef can only make a few sandwiches. If they guess wrong too often, they run out of money before finding the perfect recipe.

2. The New Idea: The "Super-Taster" AI

The authors introduce a Large Language Model (LLM) as a "Super-Taster."

The Super-Taster: This AI has read millions of cookbooks, scientific papers, and recipes. It can look at a list of ingredients and say, "I bet this combination will taste amazing," without actually cooking it.
The Cost: Asking the AI costs almost nothing (like $1). It's fast and free compared to the $100 real experiment.

3. How LABO Works: The "Gating" System

The paper's big innovation isn't just using the AI; it's knowing when to trust it and when to ignore it. They call this a "Gating Criterion."

Imagine the chef has a rule:

If the AI is confident: The chef asks the AI, "What do you think of this recipe?" If the AI says, "This looks great based on my reading," the chef doesn't cook it yet. They just write the prediction down. This lets them "taste" hundreds of recipes in their head for free, exploring the whole kitchen.
If the AI is unsure: If the AI says, "I'm not sure, this is a weird combination," or if the chef's past experience suggests the AI might be wrong, the Gate Opens. Now, the chef actually cooks the sandwich (the expensive real experiment) to get the true answer.

The Magic: The system learns to trust the AI for broad exploration (looking at the whole menu) but saves the expensive money only for the specific spots where the AI is confused or where the real answer is most uncertain.

4. The "Correction" Mechanism

The paper also mentions a clever math trick (called the KOH model).

Think of the AI's prediction as a rough sketch of the sandwich.
The "Discrepancy" is the difference between the sketch and the real taste.
LABO learns to draw a line: "The AI is usually 80% right, but here is the 20% where it gets it wrong." It uses the cheap AI sketches to get the general shape right, and only spends money to fix the specific parts where the sketch is blurry.

5. The Results: More Bang for the Buck

The paper tested this on real scientific tasks, like designing new materials for batteries or optimizing chemical reactions.

The Outcome: LABO found better recipes (solutions) faster than the old methods.
The Efficiency: It achieved the same results as the traditional method but used far fewer expensive "real" experiments. It was like finding the best sandwich by tasting 10 real ones instead of 30, because the AI helped filter out the bad ideas first.
Safety: Even if the AI sometimes gives bad advice, the "Gate" catches it. The system automatically switches back to real experiments when the AI's confidence drops, ensuring the chef doesn't waste money on a bad guess.

Summary

LABO is a smart workflow that uses a cheap, knowledgeable AI to scan the entire search space broadly. It only spends expensive money on real experiments when the AI is unsure or when the AI's guess needs to be verified. This allows scientists to find the best solutions much faster and cheaper than before, without needing to run thousands of costly physical tests.

Technical Summary: LABO - LLM-Accelerated Bayesian Optimization

Problem Statement

Scientific discovery in domains such as drug discovery, catalyst design, and molecular engineering relies heavily on formulation optimization. These tasks are characterized by expensive, time-consuming, and labor-intensive experiments, leading to severe data scarcity and limited evaluation budgets. While Bayesian Optimization (BO) is a standard strategy for navigating these complex design spaces by balancing exploration and exploitation, it faces two fundamental limitations:

The Cold-Start Problem: Data scarcity at the beginning of optimization impedes early-stage exploration.
High-Dimensional Challenges: Exploring high-dimensional design spaces remains difficult under limited experimental budgets, particularly when task-specific observations are sparse.

Although recent efforts have integrated Large Language Models (LLMs) into BO pipelines to incorporate prior knowledge (e.g., literature, protocols), existing methods typically use LLMs only for initialization or as auxiliary signal sources. They often fail to fully leverage the significantly lower evaluation cost of LLMs compared to real-world experiments, limiting their ability to support broad, low-cost exploration across the search space.

Methodology: LABO Framework

The authors propose LABO (LLM-Accelerated Bayesian Optimization), a dual-fidelity framework that systematically integrates LLM predictions with real experimental observations within a single BO loop. LABO treats LLMs as a distinct source of "LLM-fidelity" information derived from unstructured scientific priors and in-context reasoning.

1. Modeling: Kennedy–O’Hagan (KOH) Joint Gaussian Process

LABO adopts the KOH assumption to model the relationship between the real-fidelity objective $f_R(x)$ and the LLM-fidelity prediction $f_L(x)$ . The real-fidelity function is decomposed as:
$f_R(x) = \rho f_L(x) + \delta(x)$
Where:

$\rho$ is a scalar scaling factor representing the global calibration between fidelities.
$\delta(x)$ is an additive discrepancy process capturing structural differences or biases not accounted for by the LLM.
Both $f_L(x)$ and $\delta(x)$ are modeled as independent zero-mean Gaussian Processes (GPs).

This formulation allows the surrogate model to update the real-fidelity posterior using inexpensive LLM queries. If LLM predictions align well with observations, the discrepancy remains small, and the model leverages LLM priors efficiently. If predictions are unreliable, the discrepancy variance dominates, signaling the need for real-world evaluation.

2. Decision Mechanism: Discrepancy-Dominance Gating

To determine whether to query the expensive real-fidelity source or rely on the cheap LLM source, LABO employs a gating criterion based on the discrepancy dominance ratio:
$p_\Delta(x) = \frac{\sigma^2_\delta(x)}{\rho^2\sigma^2_L(x) + \sigma^2_\delta(x)}$
This ratio measures the proportion of total predictive uncertainty attributable to the discrepancy term $\delta(x)$ .

Reliable Region ( $p_\Delta(x) \le \tau$ ): If the discrepancy uncertainty is low relative to the total uncertainty, the LLM prediction is deemed reliable. LABO updates the surrogate using the LLM-fidelity GP without performing a real experiment.
Unreliable Region ( $p_\Delta(x) > \tau$ ): If the discrepancy dominates, the LLM signal is insufficient to reduce uncertainty. LABO triggers a real-fidelity experiment to update both the LLM and discrepancy models.

3. Optimization Workflow

Warm-Start: The LLM is prompted with domain knowledge to propose high-potential candidates for initial real experiments and to generate a broad set of LLM-fidelity predictions covering the search space (via Latin Hypercube Sampling).
Iterative Loop: LABO selects candidates using an acquisition function (e.g., UCB) on the composite surrogate. For each candidate, it queries the LLM, evaluates the gating criterion, and selectively performs real experiments only when necessary.

Key Contributions

Principled Framework: The proposal of LABO, a dual-fidelity BO framework that integrates LLM-derived predictions alongside real experiments using a KOH joint GP model.
Theoretical Guarantees: The establishment of cumulative regret bounds demonstrating that LABO improves sample efficiency when LLM guidance is informative. The analysis shows that the regret scales with the information capacity of a limiting real-fidelity region $X^*_R$ (where real experiments are actually needed), which is significantly smaller than the full search space, even when LLM signals are noisy or misleading.
Empirical Validation: Demonstration of LABO's practical utility across diverse scientific optimization tasks, showing it consistently finds high-performing formulations more efficiently than existing methods under fixed real-fidelity budgets.

Experimental Results

The authors evaluated LABO on six scientific optimization tasks spanning chemistry, materials science, energy systems, and nutrition (e.g., COF gas separation, Fullerene reaction yield, Flow Battery performance).

Performance: LABO consistently outperformed baselines, including Vanilla BO, LLAMBO, BOPRO, and CAKE. On high-dimensional tasks (e.g., COF, Sandwich), LABO achieved better final objective values within limited budgets. On low-dimensional tasks, it maintained stable convergence.
Sample Efficiency: The gating criterion successfully concentrated real-fidelity evaluations in high-uncertainty sub-regions while using LLM queries for broad global exploration. This resulted in a significantly reduced effective exploration region $X^*_R$ .
Robustness: Ablation studies confirmed that performance gains stem from the informative nature of LLM predictions (grounded in prior knowledge) rather than just initialization. LABO remained robust across different LLM backbones (varying sizes and reasoning capabilities) and different gating thresholds ( $\tau$ ).
Cost-Benefit: In a cost analysis on the COF task, LABO achieved the same performance level as Vanilla BO with a total cost of \ $964 (using 164 LLM queries and 8 real experiments) compared to Vanilla BO's \$ 6000 (60 real experiments), assuming a \ $1 cost per LLM query and \$ 100 per real experiment.

Significance and Claims

The paper claims that LABO offers a practical and theoretically grounded approach for integrating LLMs into scientific discovery workflows. By explicitly modeling the discrepancy between LLM predictions and reality, LABO avoids the pitfalls of blindly trusting LLMs while capitalizing on their low-cost, broad-exploration capabilities.

The authors emphasize that their approach does not require structural assumptions on the LLM's accuracy; the framework adapts to spatially varying uncertainty. The theoretical analysis suggests that LABO achieves superior sample efficiency by restricting expensive real experiments to a progressively restricted subset of the domain where LLM signals are uninformative, thereby accelerating the discovery of optimal scientific formulations without increasing experimental budgets.