General knowledge, reasoning and preference benchmarks, compared
Short answer: This page compares 6 entries: GPQA, Humanity's Last Exam, LiveBench, MMLU, LMArena (Chatbot Arena) and more. These benchmarks give a broad read on a model rather than one task: knowledge exams with known answers, contamination-limited sets, human-preference voting and multi-metric frameworks.
These benchmarks give a broad read on a model rather than one task: knowledge exams with known answers, contamination-limited sets, human-preference voting and multi-metric frameworks. They are useful for shortlisting models; test on your own task before choosing.
| Eval | Made by | What it tests | License | Last updated | Source |
|---|---|---|---|---|---|
| GPQA | NYU & collaborators | Expert-level science reasoning. | MIT | Repo last commit 2024-09-30 | github.com |
| Humanity's Last Exam | Center for AI Safety & Scale AI | Frontier academic knowledge and reasoning. | MIT | — | lastexam.ai |
| LiveBench | LiveBench team (academic) | Hard, contamination-limited tasks with verifiable ground-truth answers (no LLM judge). | — | Repo last commit 2026-10-14 | livebench.github.io |
| MMLU | Hendrycks et al. (UC Berkeley & collaborators) | World knowledge and problem solving across many school and professional subjects. | MIT | Repo last commit 2023-04-24 | github.com |
| LMArena (Chatbot Arena) | LMArena (started by LMSYS) | Which models people prefer on their own real prompts, by pairwise human votes. | — | — | lmarena.ai |
| HELM | Stanford CRFM | Standardised, multi-metric comparison of models across shared scenarios. | Apache-2.0 | Repo last commit 2026-06-05 | crfm.stanford.edu |
Head-to-head
In short
GPQA — Graduate-level science questions written by experts so that web search doesn't easily help.
Humanity's Last Exam — Very hard questions contributed by experts across many fields, made to stay difficult after older benchmarks saturated.
LiveBench — A regularly refreshed LLM benchmark with objective, auto-checkable answers across math, coding, reasoning and more, designed to limit training-data contamination.
MMLU — Multiple-choice questions across 57 subjects, from elementary mathematics and US history to computer science and law, testing broad academic and professional knowledge.
LMArena (Chatbot Arena) — A public site where people chat with two anonymous models side by side and vote for the better answer; the votes feed public leaderboards for text, coding, vision, image and video models.
HELM — Stanford's open framework and set of leaderboards for evaluating language models on many scenarios with several metrics at once, not just accuracy. Domain leaderboards include medicine (MedHELM), safety and vision-language models.
Browse the full Reasoning & knowledge category, or see other comparisons.