In materials research, reasoning tasks of different difficulty coexist, such as candidate generation, structure validation, reactivity judgment, and reading papers or patents. In such cases, one can create a common evaluation framework and policy-wise compare where to turn on search, rerank, self-consistency, and tool-use.
In practice, it is better to start with one small task. For example, in a task that ranks 100 candidate compositions, comparing several policies that combine a low-cost first-pass score and a high-cost second-stage verification under the same cost budget can quickly reveal the usefulness of the framework.