Because the difficulty of candidate materials in the materials field varies greatly, instead of using the same number of simulations or the same number of LLM reasoning steps for every composition, it may be possible to test adaptive allocation. For example, one could use the prediction variance of a surrogate model, ensemble disagreement, or an initial screening score as an uncertainty signal and allocate the next computational budget accordingly.
This approach may be especially useful in pipelines where stages with different costs are mixed, such as candidate ranking, synthesis feasibility review, experimental planning, or large-scale literature review. However, because it is necessary to first verify whether the uncertainty signal truly reflects difficulty, it is advisable to compare cost-performance curves on a small dataset.