Benchmarks for Scientific Reasoning: What a Score Establishes

Graduate-level science question sets are among the more carefully constructed benchmarks in the field, and the care went into a place that rarely gets discussed: establishing what a determined non-expert scores. That control is what turns the number into a claim.

How the questions are built

The design problem for a science benchmark is that most exam-style questions are solvable by retrieval. If a question can be answered by finding the right page, a high score measures search and memory, which is not what anybody is trying to measure.

The response, in benchmarks of the GPQA type, is a construction procedure with three steps. Domain experts — people holding or working towards doctorates in the relevant field — write questions in their own speciality. Other experts in the same field answer them, establishing that a knowledgeable person can get them right, which filters out questions that are merely ambiguous or wrong. Then skilled non-experts attempt them with unrestricted time and unrestricted web access, and questions they can solve are discarded.

That third step is the innovation and it is what makes the number interpretable. The benchmark ships with a measured human baseline for both groups, so a model score can be positioned against “determined person with a search engine” rather than against nothing. Most benchmarks do not have this, and a score without a human baseline is a number without a scale.

What a high score establishes

It is worth stating positively, because the criticisms are easier to make than the credit. A model scoring well above the non-expert baseline on questions of this kind is doing something real: it is selecting correct answers to hard, well-posed, closed-form problems across several sciences at a level that skilled people with search cannot match. Whatever else is true, the knowledge and the multi-step manipulation of it are present, and that was not the case a few model generations ago.

It is also a reasonable, if coarse, screening signal. If you are choosing a model for tasks that involve technical domain knowledge, the relative ordering on this kind of test carries some information about the ordering on your task — a weak inference, but not a worthless one, with the caveats in why benchmark results do not transfer.

Four gaps to doing science

• Multiple choice is a different task. Options can be eliminated, and elimination can reach the right answer without the reasoning that would generate it. The same questions posed as free response are harder, and the gap between the two formats is informative about how much of the score is genuine derivation.
• The questions are well-posed and have answers. Somebody verified there is a correct option. Research questions are not verified to be answerable, not verified to be well-posed, and a great deal of research skill is spent discovering that a question was wrong before anything else can proceed. That skill is not tested at all.
• No experiment, no instrument, no cost. Deciding what to measure, with what, at what precision, for what budget, and what confound will ruin it — the actual content of experimental design — has no representation in a question-answering format.
• No contradictory evidence. A benchmark question has one consistent world behind it. Real problems come with papers that disagree, a dataset with something wrong in it, and a prior result that may not replicate. Reconciling those is a large fraction of scientific work and no multiple-choice set touches it.

There is also the standing issue that any public benchmark degrades as it circulates. Questions and discussions of them end up in training corpora, and a score on a set that has been public for years is a weaker signal than the same score on a freshly written set — the subject of benchmark contamination. This is why blind, prospective evaluations carry so much more weight, which is exactly the argument made for structure prediction in the CASP format.

The other benchmark families

Question answering is only one kind of scientific benchmark, and for machine learning applied within a science the more relevant families work differently.

• Property prediction leaderboards. Predict a molecular, material or biological property from structure. Here the score is almost secondary to the split: a random split and a split by scaffold, by time or by sequence identity produce very different numbers on the same model, and only the second kind corresponds to a question anyone has. Reading these leaderboards without reading the split definition is not reading them at all.
• Protocol and procedure tasks. Given a goal, produce a laboratory procedure. Hard to score automatically and consequently rare, which is itself informative about where the measurable ground is.
• Agentic research tasks. Multi-step tasks with tools — reproduce a figure from a paper, run an analysis, debug a pipeline. Closer to real work and much harder to score reliably, with results that vary considerably across runs.

Reading a reported score

• Against which human baseline? Expert, non-expert with search, or none. The last is common and makes the number uninterpretable.
• Multiple choice or free response? They are different tasks and are frequently reported under one benchmark name.
• How old is the set? Relative to the training cutoff of the model being scored.
• What is the variance? A single run on a few hundred questions has a confidence interval wide enough to swallow most reported gaps between models.
• Was the error analysis published? Which questions were failed is far more informative than how many, and its absence usually means nobody looked.