Mixed-variable experiments stop leaning on structure alone: LLM ranking priors plus structural smoothing reach near-optimal candidates faster on six chemistry and materials tasks

In mixed-variable scientific optimization, similarity among discrete choices is no longer set by structural descriptors alone; it is also set by task-relevant ordering relations. Earlier Bayesian optimization methods typically define categorical similarity from molecular fingerprints, chemical group encodings, and other structural descriptors, implicitly assuming that structurally close candidates behave similarly on the target task. Across six optimization tasks in pharmaceuticals, chemistry, and materials science, TagBO shows consistently lower Step@95% values under the same evaluation budget, meaning it approaches near-optimal solutions with fewer evaluations.

The method has a large language model first list the key factors likely to influence the objective, then rank candidates under each factor, and treats those ranking vectors as coarse task-aware relational signals. Those ranking signals are not used directly as predictive scores; they are combined with physicochemical descriptor similarity from external databases such as PubChem, and a refined representation is learned through a ranking-preservation loss and a graph-consistency loss. The two objectives are balanced adaptively by the multi-gradient descent algorithm rather than a manually fixed weight, and Gaussian process Bayesian optimization with feasibility constraints then runs over the concatenated discrete and continuous representation space.

Ablations show the two information sources are complementary: the LLM-only variant is strong early but saturates later, while the structural-only variant stays weaker throughout. Only combining coarse relational guidance with fine-grained structural smoothness yields both strong early exploration and improved later refinement. Comparing the full method with two reduced variants across six datasets, the full method generally reaches strong objective values in fewer optimization steps while achieving comparable or better final performance.

The report covers mixed-variable optimization in chemistry and materials formulation: discrete choices plus continuous parameters, with costly evaluations. The intended readers are researchers and engineers working on experimental design, Bayesian optimization, and self-driving laboratories. Real-world validation is limited to a thermally conductive adhesive formulation with 34 physical evaluations, each taking about three days. The authors propose retrieval-augmented generation and domain-specific knowledge integration as future work to address weak LLM priors.

Effectiveness depends on the quality of the task-aware relational priors extracted from the LLM; in domains where relevant scientific knowledge is sparse, highly specialized, or poorly represented in the pretraining corpus, the induced ordinal relations may become noisy or misaligned with the downstream objective. The theoretical analysis addresses representation-level surrogate behavior rather than cumulative regret bounds for the full sequential optimization process.