These suites test mathematical reasoning, research-level science, biology, physics, and scientific coding. Many are built to resist contamination, and top scores on some are still low.
What do hard science and reasoning benchmarks tell us?
Use them to find the limits of current models and to compare reasoning settings. They do not show that a model can do research alone. Check whether tools were on, answers are machine-checked, tasks are private, and runs were repeated.
21
Benchmarks in this domain
Sources first · alphabetical collection order
Runnability describes access to a usable repository, dataset, or harness—not whether setup is easy.
AIME 2026
specializedrunnable
2026-11-03
Tests
Tests math problem solving on the 30 problems of the 2026 AIME (American Invitational Mathematics Examination). Each answer is an integer.
Best for
A dated, difficult math signal with explicit versioning.
Tests LLMs (large language models) on research-level physics problems that physicists wrote from their own unpublished work. Answers can be checked by machine.
Gives the model several hundred unpublished math problems written by mathematicians. Tiers 1-3 go from undergraduate to advanced graduate level. Tier 4 is research level.
Best for
Comparing the strongest models on hard math that is not in public training data.
Tests science knowledge with multiple-choice questions in biology, physics, and chemistry. PhD experts reach 65 percent, and skilled non-experts with web access reach 34 percent.
Gives the model the problems from the February 2026 HMMT (Harvard-MIT Mathematics Tournament) competition. Checks each final answer against a gold answer.
Best for
A fresh math signal that complements AIME-style evaluation.
Gives the model 2,500 expert-written closed-ended questions across many academic subjects, some with images. Checks multiple-choice and short answers automatically.
Tests whether an AI system can do practical biology research tasks. Tasks cover literature, figures, databases, protocols, and DNA and protein sequences. LABBench2 puts similar tasks in more realistic contexts.
Gives the model expert-written life-science research tasks with a grading rubric for each task. Tests realistic research work, not only biology question answering.
Best for
Evaluating whether systems can support realistic research work beyond biology QA.
Gives the model new questions from recent sources such as math competitions, arXiv papers, and news. Covers math, coding, reasoning, language, instruction following, and data analysis.
Best for
General model signal that resists stale benchmark gaming
Gives the model multiple-choice questions across academic and professional subjects, with ten answer options. It adds harder reasoning questions and removes noisy MMLU questions.
Tests how well a protein structure model learns from a small fixed training set. Public data has 10,000 training chains and 1,000 validation chains. Ranking uses hidden validation chains.
Best for
Specialist life-science model comparisons with a public result feed.
Gives AI agents about 200 real research tasks with step-by-step verification in an interactive environment. Tests whether agents can do scientific data analysis and open-ended research.
Best for
Comparing research-agent reliability, cost, and task-level scientific contribution.
Gives the model unusual hand-made machine learning tasks with unfamiliar data. In v3, an agent must build ML (machine learning) pipelines with limited data, unclear goals, or little feedback.