AIME 2026
- Tests
- Advanced mathematical problem solving on the 2026 AIME contest set.
- Best for
- A dated, difficult math signal with explicit versioning.
- Classification
- benchmark · emerging · mixed
AI evaluation domain · 19 source-backed records
Science, Math, and Reasoning Benchmarks
These suites probe mathematical reasoning, research-level science, biology, physics, and scientific coding. Many are intentionally frontier or contamination-resistant and may have very low scores.
Decision guidance
What do hard science and reasoning benchmarks tell us?
Use them to locate capability boundaries and compare reasoning setups, not to claim research autonomy. Inspect whether tools were enabled, answers are machine-verifiable, tasks were private, and repeated runs were used.
19
Benchmarks in this domain
Sources first · alphabetical collection order
Runnability describes access to a usable repository, dataset, or harness—not whether setup is easy.
2026-07-15
AIRS-Bench
- Tests
- End-to-end ML research ability on tasks derived from state-of-the-art papers.
- Best for
- Comparing AI research agents on executable experiment outcomes.
- Classification
- benchmark · emerging · open
ARC-AGI-2
- Tests
- Abstract visual reasoning tasks focused on generalization from few examples.
- Best for
- Abstract reasoning and generalization stress testing
- Classification
- benchmark · current · mixed
BBH / BIG-Bench Hard
- Tests
- Hard tasks from BIG-bench covering symbolic, logical, and multi-step reasoning.
- Best for
- Legacy reasoning regression checks
- Classification
- benchmark · legacy · open
CritPt
- Tests
- Research-level physics tasks from modern physics areas.
- Best for
- Niche frontier physics reasoning
- Classification
- benchmark · current · open
GPQA / GPQA Diamond
- Tests
- Graduate-level science questions designed by domain experts and difficult for non-experts.
- Best for
- Expert science reasoning
- Classification
- benchmark · current · open
GSM8K
- Tests
- Multi-step arithmetic reasoning on grade-school word problems.
- Best for
- Historical baselines and regression checks; it is increasingly saturated for frontier-model selection.
- Classification
- benchmark · legacy · open
HMMT February 2026
- Tests
- Advanced competition mathematics on the February 2026 HMMT set.
- Best for
- A fresh math signal that complements AIME-style evaluation.
- Classification
- benchmark · emerging · mixed
Humanity's Last Exam
- Tests
- Expert-level closed-ended multimodal academic questions intended for frontier model differentiation.
- Best for
- Hard frontier academic reasoning
- Classification
- benchmark · current · open
LAB-Bench / LABBench2
- Tests
- Biology research tasks covering literature, protocols, databases, DNA/protein sequences, and lab reasoning; LABBench2 adds a newer dataset and harness.
- Best for
- Biology research assistants
- Classification
- benchmark · current · open
LifeSciBench
- Tests
- Expert-written, expert-reviewed tasks grounded in practical life-science research.
- Best for
- Evaluating whether systems can support realistic research work beyond biology QA.
- Classification
- benchmark · emerging · mixed
LiveBench
- Tests
- Fresh, contamination-conscious questions with objective scoring across reasoning, math, coding, language, and data analysis.
- Best for
- General model signal that resists stale benchmark gaming
- Classification
- benchmark · current · open
MATH / AIME-style evals
- Tests
- Competitive math problem solving and formal reasoning tasks.
- Best for
- Mathematical reasoning and launch-card comparison
- Classification
- benchmark · current · open
MLE-bench
- Tests
- Machine-learning engineering tasks drawn from Kaggle-style competitions.
- Best for
- ML engineering agents
- Classification
- agent benchmark · current · open
MMLU-Pro
- Tests
- Harder multiple-choice academic and professional knowledge benchmark derived from MMLU with more options.
- Best for
- Broad expert knowledge screening
- Classification
- benchmark · current · open
NanoFold Public
- Tests
- Scientific-model performance on public protein-folding tasks.
- Best for
- Specialist life-science model comparisons with a public result feed.
- Classification
- benchmark · emerging · open
Pencil Puzzle Bench
- Tests
- Deterministically verifiable constraint-satisfaction puzzle tasks.
- Best for
- Reasoning without LLM judges
- Classification
- benchmark · current · open
SciAgentArena
- Tests
- Scientific agents on about 200 stepwise-verified tasks across five biomedical domains.
- Best for
- Comparing research-agent reliability, cost, and task-level scientific contribution.
- Classification
- benchmark · emerging · open
WeirdML
- Tests
- Unusual ML tasks designed to reward actual understanding over rote benchmark skill.
- Best for
- Anti-gaming ML reasoning
- Classification
- benchmark · current · mixed