On 27 October 2026, the top 10 LLM evaluation benchmarks, ranked by citations to the paper that introduced each one, were HumanEval, GSM8K, MMLU, MATH, MBPP, SWE-bench, GPQA, LiveCodeBench, MMLU-Pro and BIG-Bench Hard. HumanEval leads with 12,391 citations in Semantic Scholar.
This list does not start from a list of famous benchmarks. It starts from what frontier labs actually report. A benchmark qualifies only if it is named in at least 3 of 6 technical reports for 2025's leading language models — Qwen3, DeepSeek-R1, Gemini 2.5, Gemma 3, gpt-oss and DeepSeek-V3. That rule is what keeps a widely cited but no longer used benchmark off the list, and every figure below links to its record.
What are the top 10 LLM evaluation benchmarks?
| Rank | Benchmark | What it tests | Citations | Influential | Used in reports | Introduced |
|---|---|---|---|---|---|---|
| 1 | HumanEval | Python code from docstrings | 12,391 | 1,853 | 4 of 6 | 2021 |
| 2 | GSM8K | Grade-school maths word problems | 11,690 | 2,599 | 3 of 6 | 2021 |
| 3 | MMLU | Knowledge across 57 subjects | 10,245 | 1,944 | 6 of 6 | 2020 |
| 4 | MATH | Competition mathematics | 7,082 | 1,598 | 4 of 6 | 2021 |
| 5 | MBPP | Entry-level programming tasks | 4,915 | 768 | 3 of 6 | 2021 |
| 6 | SWE-bench | Resolving real GitHub issues | 4,564 | 756 | 4 of 6 | 2023 |
| 7 | GPQA | Expert-written science questions | 3,801 | 696 | 6 of 6 | 2023 |
| 8 | LiveCodeBench | Fresh contest coding problems | 2,503 | 380 | 5 of 6 | 2024 |
| 9 | MMLU-Pro | A harder, more robust MMLU | 2,478 | 270 | 4 of 6 | 2024 |
| 10 | BIG-Bench Hard | The 23 hardest BIG-Bench tasks | 2,355 | 329 | 3 of 6 | 2022 |
Citation counts were read from Semantic Scholar on 27 October 2026 and rise every week. "Used in reports" is how many of the six frontier technical reports name the benchmark.
Where can you verify each benchmark's numbers?
Each benchmark links to the arXiv page of its introducing paper and to that paper's Semantic Scholar record.
- HumanEval — arXiv · Semantic Scholar
- GSM8K — arXiv · Semantic Scholar
- MMLU — arXiv · Semantic Scholar
- MATH — arXiv · Semantic Scholar
- MBPP — arXiv · Semantic Scholar
- SWE-bench — arXiv · Semantic Scholar
- GPQA — arXiv · Semantic Scholar
- LiveCodeBench — arXiv · Semantic Scholar
- MMLU-Pro — arXiv · Semantic Scholar
- BIG-Bench Hard — arXiv · Semantic Scholar
How were these benchmarks selected?
A ranking of benchmarks by citations alone would reward age and fame. This one adds a test of current use first.
- Read six frontier technical reports in full: Qwen3, DeepSeek-R1, Gemini 2.5, Gemma 3, the gpt-oss model card and DeepSeek-V3.
- Count which benchmarks each report names, searching for 51 benchmark names and keeping variants apart — MMLU is not MMLU-Pro, and BIG-bench is not BIG-Bench Hard.
- Keep benchmarks named in at least 3 of the 6. That leaves 20.
- Drop those with no introducing paper, since a citation count needs a paper. AIME and Codeforces are competitions; Aider's benchmark is a leaderboard. 17 remain.
- Rank the 17 by Semantic Scholar citations to the introducing paper.
A completeness check scanned all six reports for benchmark-shaped names outside the search list. It found one: MMMLU, a multilingual translation of MMLU, named in three reports — but with no introducing paper to rank it by.
Why is BIG-bench not on the list?
Because none of the six reports uses it. Its introducing paper has more citations than three of the ten above, but the reports use its descendants instead: BIG-Bench Hard, a 23-task subset, and BIG-Bench Extra Hard. A first pass counted BIG-bench in three reports; every one of those matches turned out to be the title of the BIG-Bench Hard paper in a reference list, or the name BIG-Bench Extra Hard. That is the difference between being famous and being used.
Which benchmarks do frontier labs actually rely on?
Citations measure academic attention. The report count measures what model builders choose to show. The two rank the same benchmarks very differently:
| Benchmark | Used in reports | Citation rank |
|---|---|---|
| MMLU | 6 of 6 | 3 |
| GPQA | 6 of 6 | 7 |
| LiveCodeBench | 5 of 6 | 8 |
| SimpleQA | 5 of 6 | 16 |
| GSM8K | 3 of 6 | 2 |
| MBPP | 3 of 6 | 5 |
- MMLU and GPQA are the only benchmarks in all six reports. MMLU is six years old and still universal; GPQA, with 448 questions, is the hard-science test every lab reports.
- SimpleQA is in five of the six reports but has only 383 citations, sixteenth of 17. Introduced in November 2024, it measures short-form factual accuracy and has been adopted by labs far faster than it has been cited by papers.
- GSM8K is second by citations but in only three reports. The three that leave it out are the reasoning-focused ones — DeepSeek-R1, Gemini 2.5 and gpt-oss.
Why do some benchmarks have so many citations?
Partly because their papers are about more than the benchmark. Three of the top five introducing papers did more than release a dataset, and a citation to any part of a paper counts the same:
- HumanEval was released in the Codex paper, which introduced a code model as well as the test set. Its citations include every paper that cites Codex.
- GSM8K came with a paper on training verifiers to check maths solutions.
- MBPP appeared in a wider study of program synthesis with large language models.
A citation count for these is an upper bound on attention to the benchmark itself. The influential-citation column, which counts papers that build on the work substantially, is a better guide, and by it GSM8K leads with 2,599, ahead of MMLU's 1,944 and HumanEval's 1,853.
What does each benchmark test?
1. HumanEval — 12,391 citations
Functional correctness of Python programs synthesised from docstrings, introduced with the Codex model in July 2021.
2. GSM8K — 11,690 citations
8.5K grade-school maths word problems that need multi-step reasoning.
3. MMLU — 10,245 citations
Multitask accuracy across 57 subjects, from elementary mathematics and US history to computer science and law. Named in all six reports.
4. MATH — 7,082 citations
12,500 challenging competition mathematics problems.
5. MBPP — 4,915 citations
974 programming tasks designed to be solvable by entry-level programmers.
6. SWE-bench — 4,564 citations
2,294 software engineering problems drawn from real GitHub issues and pull requests across 12 popular Python repositories.
7. GPQA — 3,801 citations
448 multiple-choice questions in biology, physics and chemistry, written by domain experts and "Google-proof": skilled non-experts with web access reach only 34%.
8. LiveCodeBench — 2,503 citations
Coding problems collected continuously from contests on LeetCode, AtCoder and Codeforces, so that new problems post-date a model's training.
9. MMLU-Pro — 2,478 citations
A harder, more prompt-robust successor to MMLU, which its authors report cuts accuracy by 16% to 33% compared with the original.
10. BIG-Bench Hard — 2,355 citations
A suite of 23 challenging BIG-Bench tasks on which language models had fallen short of average human performance.
Which benchmark just missed the top ten?
DROP, a reading-comprehension benchmark requiring discrete reasoning over paragraphs, at 1,444 citations — named in three of the six reports. Behind it are IFEval (1,312), C-Eval (986), MGSM (704), Arena-Hard (584), SimpleQA (383) and MMLU-Redux (218).
What are the limits of this ranking?
- Six reports are a sample. They cover the leading open and closed models of 2025, but a different set would move the threshold for a few benchmarks.
- Naming is not the same as headline use. A benchmark counts if a report names it, which includes a comparison table, an appendix or a reference.
- Citations conflate benchmark and paper, as above, for HumanEval, GSM8K and MBPP in particular.
- Competitions are excluded by design. AIME and Codeforces are among the most reported evaluations and have no introducing paper to count citations for.
How is this different from being cited by an AI search engine?
Academic citations count how often researchers reference a paper. AI-search citation is a different measure: whether an assistant names or links your page when someone asks a question. That is tracked across a set of real queries, by which sources and brands get cited. If that is the citation you care about, start with what AI visibility is or how to track brand mentions in AI search.
Frequently asked questions
Ranked by citations to their introducing papers on 27 October 2026, among benchmarks named in at least 3 of 6 frontier technical reports: HumanEval (12,391), GSM8K (11,690), MMLU (10,245), MATH (7,082), MBPP (4,915), SWE-bench (4,564), GPQA (3,801), LiveCodeBench (2,503), MMLU-Pro (2,478) and BIG-Bench Hard (2,355).
MMLU and GPQA are the only two named in all six of the frontier technical reports read for this ranking: Qwen3, DeepSeek-R1, Gemini 2.5, Gemma 3, gpt-oss and DeepSeek-V3. LiveCodeBench and SimpleQA are each in five.
HumanEval, with 12,391 Semantic Scholar citations. It was introduced in the Codex paper, so its count also includes citations to the Codex model itself.
None of the six frontier reports uses it. They use its successors instead: BIG-Bench Hard, a 23-task subset, and BIG-Bench Extra Hard. A benchmark only qualifies here if at least three of the six reports name it.
They are competitions rather than benchmarks with an introducing paper, so there is no citation count to rank them by. Both are named in several of the reports.
Academic citations count researchers referencing a paper. AI search visibility is whether an AI assistant names or links your page when someone asks a question, tracked across a set of real queries.
The short version
On 27 October 2026 the most cited LLM evaluation benchmarks still in use by frontier labs were HumanEval, GSM8K and MMLU, followed by MATH, MBPP, SWE-bench, GPQA, LiveCodeBench, MMLU-Pro and BIG-Bench Hard. The list only includes benchmarks named in at least three of six 2025 frontier technical reports, which is why BIG-bench is missing despite its fame. MMLU and GPQA are the only two every report uses.