Benchmarks for Scientific Reasoning: What a Score Establishes

Benchmarks in scientific reasoning are meticulously designed to measure a model's ability to answer complex, well-posed scientific questions. Most exam-style questions can be solved by retrieval, but these benchmarks aim to go beyond that. Domain experts write questions in their specialty, which are then verified by other experts to ensure they are solvable and not just ambiguous or incorrect.

Finally, non-experts with unrestricted time and web access attempt the questions, and only those solvable by a determined non-expert are retained. This process establishes a human baseline for scores, allowing models to be evaluated against "a determined person with a search engine" rather than against nothing.

A high score on these benchmarks indicates that the model is capable of selecting correct answers to difficult, well-posed problems across multiple scientific domains, a level that skilled humans with search engines cannot match. This demonstrates that the model possesses knowledge and the ability to manipulate it through multi-step reasoning. The scores also serve as a reasonable screening signal for choosing models to tackle tasks involving technical domain knowledge.

However, there are limitations to what these benchmarks can measure. Multiple-choice questions allow for elimination of options, which can lead to correct answers without the reasoning required to arrive at them. Free-response questions, on the other hand, are harder and provide a more accurate representation of the model's genuine derivation.

Additionally, the questions are well-posed and have known answers, but the research questions themselves may not be well-posed or answerable, as much of the effort in research goes into discovering flaws in questions before proceeding.

Experimental design, including deciding what to measure, how to measure it, and how to handle confounding factors, is absent from the benchmark format. Real-world problems often come with contradictory evidence, conflicting data, and non-replicable results, which are not accounted for in multiple-choice formats. Furthermore, public benchmarks can degrade over time as they become incorporated into training corpora, reducing the interpretability of scores from older sets.

Other families of scientific benchmarks include property prediction leaderboards, protocol and procedure tasks, and agentic research tasks. These benchmarks differ in their scoring methods, difficulty, and relevance to real-world applications. When evaluating a reported score, it is crucial to consider which human baseline was used (expert, non-expert with search, or none), whether the questions are multiple-choice or free response, the age of the benchmark set relative to the model's training data, the variance of the results, and whether an error analysis was published.