ExLLM: Experience-Enhanced LLM Optimization for Molecular Design and Beyond

TL;DR

ExLLM reframes molecular design as an optimization task guided by an LLM, introducing three lightweight yet powerful mechanisms: a compact evolving experience to distill non-redundant cues, a $k$-offspring exploration strategy to widen search per query, and a unified feedback adapter to harmonize heterogeneous signals. With a fixed evaluation budget, ExLLM achieves state-of-the-art PMO performance and strong cross-domain results, including circle packing and stellarator design, while maintaining efficiency across multiple LLM backbones. The framework transfers with a simple task-template and evaluation functions, and demonstrates robust exploration, scalable memory usage, and practical applicability to diverse scientific design problems. Together, these contributions offer a practical, transfer-ready optimizer for large discrete spaces in chemistry and beyond.

Abstract

Molecular design involves an enormous and irregular search space, where traditional optimizers such as Bayesian optimization, genetic algorithms, and generative models struggle to leverage expert knowledge or handle complex feedback. Recently, LLMs have been used as optimizers, achieving promising results on benchmarks such as PMO. However, existing approaches rely only on prompting or extra training, without mechanisms to handle complex feedback or maintain scalable memory. In particular, the common practice of appending or summarizing experiences at every query leads to redundancy, degraded exploration, and ultimately poor final outcomes under large-scale iterative search. We introduce ExLLM (Experience-Enhanced LLM optimization), an LLM-as-optimizer framework with three components: (1) a compact, evolving experience snippet tailored to large discrete spaces that distills non-redundant cues and improves convergence at low cost; (2) a simple yet effective k-offspring scheme that widens exploration per call and reduces orchestration cost; and (3) a lightweight feedback adapter that normalizes objectives for selection while formatting constraints and expert hints for iteration. ExLLM sets new state-of-the-art results on PMO and generalizes strongly in our setup, it sets records on circle packing and stellarator design, and yields consistent gains across additional domains requiring only a task-description template and evaluation functions to transfer.

Figures (20)

• Figure 1: ExLLM requires only two user inputs: a task-description template and evaluation functions, and operates without domain-specific training or tuning. The figure shows how a single unified workflow transfers seamlessly across different discrete problem optimization classes (chemical, geometric, combinatorial, physical, and code-generation tasks), achieving state-of-the-art or record-level performance under domain shift.
• Figure 2: Overall framework of ExLLM. The process begins with an initialized population, followed by LLM-based $k$-offspring generation, evaluation and feedback aggregation, hybrid selection (fitness + Pareto), and experience update. These steps repeat until the evaluation budget is exhausted.
• Figure 3: Best layouts found by our optimizer: (left) $n=26$, (right) $n=32$.
• Figure 4: A visualization of our best solution for problem 2.
• Figure 5: Comparison of the offshore jacket structure before and after optimization. Through structural optimization, the member dimensions were adjusted to satisfy all safety and performance constraints (e.g., stress, displacement), resulting in a total weight reduction of 93%.