What it measures

College-level textbook problems in mathematics, chemistry, and physics requiring open-ended numerical derivations.

Reading the measurement

Ask whether the evaluation rewards recalling a fact, constructing a model, running an experiment or supporting a discovery. Those tasks require different evidence. For research claims, inspect reproducibility, uncertainty and what was actually checked.

Published evaluation

Accuracy averaged over textbook problems across chemistry, physics and mathematics. Section labels retain zero/few-shot prompting, chain of thought and tool settings.

Where this fits in an evaluation suite

A correct multiple-choice answer and a supported scientific contribution demand different evidence. Check whether the score rewards a proposal, an executed experiment, or a completed research target.

Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.

A reproducible comparison

Accuracy averaged over textbook problems across chemistry, physics and mathematics. Section labels retain zero/few-shot prompting, chain of thought and tool settings.

Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.

PRIMARY EVIDENCE & FURTHER READING

  1. SciBench: Evaluating College-Level Scientific Problem-Solving Abilities of Large Language Models — published result table ↗

    Wang, Xiaoxuan, Hu, Ziniu, Lu, Pan, Zhu, Yanqiao, Zhang, Jieyu, Subramaniam, Satyen et al. · 2023

    All 10 authors

    Wang, Xiaoxuan, Hu, Ziniu, Lu, Pan, Zhu, Yanqiao, Zhang, Jieyu, Subramaniam, Satyen, Loomba, Arjun R., Zhang, Shichang, Sun, Yizhou, Wang, Wei

Reviewed 2026-10-27. Explanation and interpretation by RLTasks. Benchmark ownership remains with its original authors.