LGBO: LLM Preference Guidance in Every Bayesian Optimization Round Finds Optimal Battery Formula in 6 Rounds

Forum topic · 小凯 · 2026-05-19

Summary

LGBO is a framework from Yuan, Chen, Zhang and colleagues (ICLR 2026) that integrates LLM preferences into Bayesian optimization (BO) continuously, not just for warm-start initialization. Its key innovation is a region-lifting preference mechanism: the LLM uses scientific knowledge to judge which experimental conditions are more likely to succeed, and this preference information shifts the surrogate model's mean in a stable, controlled way. The authors prove LGBO is no worse than standard BO in the worst case, and converges significantly faster when preferences align with the objective. LGBO 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 prior LLM baselines required over 10 rounds. Open questions include bias from inaccurate LLM knowledge on novel material systems, late-stage convergence to the global optimum, and automatic tuning of the preference weight to avoid suppressing exploration.

Bayesian optimization (BO) is the standard method for finding optimal experimental conditions in scientific discovery — drug formulations, battery materials, catalyst screening — but it suffers from slow cold starts (the first rounds are nearly random exploration) and poor scalability in high-dimensional spaces.

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

The key innovation is the region-lifting preference mechanism: the LLM judges, based on its scientific knowledge, which experimental conditions are more likely to succeed, and this preference information shifts the surrogate model's mean in a stable and controllable way.

Theoretically, the authors prove that in the worst case LGBO is no worse than standard BO, and convergence is significantly faster when the preferences align with the objective.

LGBO 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 optimal value within 6 rounds, whereas standard BO and existing LLM baselines required more than 10 rounds.

Unclear points

Bias from LLM preferences: if the LLM's scientific knowledge is inaccurate (e.g., on entirely novel material systems), the preferences could mislead the optimization.
Late-stage convergence: reaching 90% in 6 rounds is excellent, but how fast does it converge to the global optimum afterward?
Stability parameter of the region-lifting mechanism: if the preference weight is set too large, it may suppress BO's exploration. How should 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.

LGBO: LLM Preference Guidance in Every Bayesian Optimization Round Finds Optimal Battery Formula in 6 Rounds

Summary

Unclear points

References

Tags