Unleashing LLMs in Bayesian Optimization: Preference-Guided Framework for Scientific Discovery

2026-05-18 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Research Methodology & Innovation · Depth: Expert, medium

Summary

Xinzhe Yuan et al. propose LLM-Guided Bayesian Optimization (LGBO), a novel framework that integrates large language models (LLMs) into the Bayesian Optimization (BO) loop to accelerate scientific discovery. Traditional BO struggles with slow cold-starts and poor scalability in high-dimensional problems. LGBO addresses these issues by introducing a "region-lifted preference mechanism" that embeds LLM-driven semantic reasoning into every iteration, stably shifting the surrogate mean. Unlike prior methods that use LLMs only for warm-start or candidate generation, LGBO continuously leverages LLM preferences. Theoretically, LGBO performs no worse than standard BO in the worst case but achieves significantly faster convergence when preferences align with the objective. Empirically, LGBO consistently outperforms existing methods across diverse benchmarks in physics, chemistry, biology, and materials science. Notably, in a wet-lab optimization of Fe-Cr battery electrolytes, LGBO achieved 90% of the best observed value within 6 iterations, compared to over 10 iterations for standard BO and other LLM-augmented baselines.

Key takeaway

For AI Scientists and Research Scientists working on scientific discovery with costly experiments, LGBO offers a significant advantage by accelerating optimization. You should consider implementing LGBO to overcome the cold-start and scalability limitations of traditional Bayesian Optimization, especially in high-dimensional settings. This framework can achieve faster convergence and reduce the number of experimental iterations required, as demonstrated by achieving 90% of the best value in Fe-Cr battery electrolytes within 6 iterations.

Key insights

LGBO integrates LLM semantic reasoning into every BO iteration via a preference mechanism, accelerating scientific discovery.

Principles

• LLM preferences can stably shift BO surrogate means.
• Continuous LLM integration outperforms warm-start only.
• Preference alignment accelerates BO convergence.

Method

LGBO embeds LLM-driven preferences into each BO iteration using a region-lifted mechanism, which shifts the surrogate mean to guide optimization.

In practice

• Apply LGBO for high-dimensional scientific optimization.
• Use LGBO to accelerate Fe-Cr battery electrolyte discovery.
• Integrate LLMs beyond warm-starts for BO.

Topics

• Bayesian Optimization
• Large Language Models
• Scientific Discovery
• Preference-Guided Optimization
• High-Dimensional Optimization

Code references

• haoxuan-unt2024/llm4innovation

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Automated Kernel Discovery Towards Understanding High-dimensional Bayesian Optimization — LLM-driven kernel discovery via a two-stage process and LOO-CRPS significantly improves high-dimensional Bayesian optimization.
• MetaEvo: A Meta-Optimization Framework for Experience-Driven Agent Evolution — MetaEvo enables LLM agents to continually improve reasoning by optimizing how they learn principles from experience, not just what they store.
• Training LLMs at test time for Scientific Discovery — Test-time training with an Entropic objective enables LLMs to self-improve and discover novel solutions by prioritizing high-reward outliers.
• Mind the Gap: Structure-Aware Consistency in Preference Learning — Standard preference learning surrogates are theoretically inconsistent for LLM alignment, requiring structure-aware margin adaptation.
• Time-Aware Latent Space Bayesian Optimization — TALBO adapts latent-space Bayesian optimization to dynamic objectives by integrating time into both the surrogate and generative representation.
• SMCEvolve: Principled Scientific Discovery via Sequential Monte Carlo Evolution — SMCEvolve grounds LLM-driven program evolution in a rigorous probabilistic framework with convergence guarantees.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.