LGBO: LLM Preference Guidance in Every Round of Bayesian Optimization — Optimal Battery Recipe in 6 Rounds

Forum topic · 小凯 · 2026-05-19

Summary

Bayesian optimization (BO) is the standard method for finding optimal experimental conditions in scientific discovery, but it suffers from slow cold starts and poor scalability in high-dimensional spaces. LGBO, presented by Yuan, Chen, Zhang and colleagues (ICLR 2026), is the first framework to continuously embed LLM preferences into the BO loop — not just for warm-start initialization, but injecting LLM semantic reasoning at every optimization round. Its key innovation is a 'regional uplift preference mechanism': the LLM judges which experimental conditions are more likely to succeed based on scientific knowledge, shifting the surrogate model's mean in a stable, controlled way. Theoretically, LGBO is provably no worse than standard BO in the worst case, and converges significantly faster when preferences align with the objective. It consistently outperforms existing methods on dry-run benchmarks across physics, chemistry, biology, and materials science. In wet-lab experiments on Fe-Cr battery electrolytes, LGBO reached 90% of the optimum within 6 rounds, while standard BO and existing LLM baselines required more than 10 rounds.

Bayesian optimization (BO) is the standard approach for finding optimal experimental conditions in scientific discovery — new drug formulations, battery materials, catalyst screening — but it has two well-known problems: a slow cold start (the first rounds are nearly random exploration) and poor scalability in high-dimensional spaces.

LGBO, proposed by Yuan, Chen, Zhang and colleagues (ICLR 2026), is the first framework to continuously embed LLM preferences into the BO loop — not merely for warm-start initialization, but injecting LLM semantic reasoning into every round of optimization.

Key innovation: the regional uplift preference mechanism

The LLM uses its scientific knowledge to judge which experimental conditions are more likely to succeed.
The preference information shifts the surrogate model's mean in a stable and controllable way.
Theoretical guarantees: in the worst case, LGBO is no worse than standard BO; when preferences align with the true objective, convergence is significantly faster.

Results

Consistently outperforms existing methods on dry-run benchmarks across physics, chemistry, biology, and materials science.
In wet-lab experiments on Fe-Cr battery electrolytes, LGBO reached 90% of the optimum value within 6 rounds, while standard BO and existing LLM-based baselines needed more than 10 rounds.

Open questions

Bias from LLM preferences: if the LLM's scientific knowledge is inaccurate (e.g., for entirely novel material systems), the preferences could mislead the optimization.
Post-6-round convergence: reaching 90% in 6 rounds is impressive, but how fast does it subsequently converge to the global optimum?
Stability parameter of the regional uplift mechanism: if the preference weight is set too large, it may suppress BO's exploration. How can it be determined automatically?

References

1. Yuan, X., Chen, Z., Zhang, J., et al. (2026). *Unleashing LLMs in Bayesian Optimization: Preference-Guided Framework for Scientific Discovery*. arXiv:2605.17976 [cs.AI]. 2. Frazier, P. I. (2018). *A Tutorial on Bayesian Optimization*. arXiv. 3. Jablonka, K. M., et al. (2025). *Large Language Models for Scientific Discovery*. Nature Computational Science.