Benchmarks

Standardized benchmarks used to measure AI system capabilities across reasoning, language, perception, and more.

25 benchmarks

Benchmark	Description	Leaderboard
ARC-AGI-2 arc-agi-2	Abstract reasoning benchmark testing fluid intelligence via novel grid puzzles.	View →
AudioSet audioset	Sound classification across 632 audio categories; mean average precision metric.	View →
BEHAVIOR-1K behavior-1k	Long-horizon robot task planning and execution in novel environments.	View →
BIG-Bench Hard big-bench-hard	Hard subset of BIG-Bench testing genuine few-shot generalization.	View →
BigToM bigtom	Theory of mind at scale; tests higher-order belief reasoning.	View →
CLadder cladder	Causal reasoning benchmark grounded in Pearl's causal hierarchy.	View →
CULTURAL-BENCH cultural-bench	Cultural knowledge and reasoning across 45 countries, 1,000+ questions.	View →
Chatbot Arena (LMSYS) chatbot-arena	Human preference ELO ranking across language generation and conversation tasks.	View →
EmpatheticDialogues empathetic-dialogues	Empathetic response generation in open-domain conversation.	View →
FLORES-200 flores-200	Translation quality benchmark across 200 languages.	View →
FOLIO folio	First-order logic reasoning benchmark in natural language.	View →
FrontierMath frontiermath	Novel mathematical problem solving requiring genuine reasoning, not recall.	View →
GAIA gaia	Multi-step real-world task completion with verifiable answers; levels 1–3.	View →
GPQA Diamond gpqa-diamond	Graduate-level reasoning in physics, biology, and chemistry. Resists pattern matching.	View →
ImageNet-1K imagenet	Image classification benchmark; 1,000 object categories, top-1 accuracy metric.	View →
LibriSpeech librispeech	Speech recognition benchmark on clean and noisy audio; word error rate metric.	View →
MMMU mmmu	Massive multi-discipline multimodal understanding across text and images.	View →
ManiSkill maniskill	Robotic manipulation benchmark across dexterous and contact-rich tasks.	View →
ManiSkill-ViTac maniskill-vitac	Tactile and vision-tactile fusion manipulation challenge.	View →
PARTNR partnr	Human-robot collaboration across 100,000 household tasks in simulation.	View →
Split-CIFAR-100 split-cifar-100	Continual learning benchmark measuring catastrophic forgetting across tasks.	View →
SuperGLUE superglue	Multi-task NLU benchmark testing reading comprehension, coreference, and entailment.	View →
VTAB-1k vtab-1k	Transfer learning across 19 diverse visual tasks from 1,000 examples.	View →
VoxCeleb1 voxceleb	Speaker verification benchmark; equal error rate metric.	View →
WinoGrande winogrande	Commonsense reasoning via pronoun resolution; resists surface shortcuts.	View →