Skip to content

Benchmark glossary

Research cut-off date: 2026-07-22. Status: written. Definitional entries only. No scores appear here; results live in data/benchmarks.csv with their evaluation conditions attached.

Benchmark What it tests Primary use
GPQA Diamond Graduate-level, expert-written science questions Scientific knowledge and reasoning
Humanity's Last Exam Difficult multidisciplinary academic questions Broad frontier reasoning
ARC-AGI-2 and ARC-AGI-3 Abstract reasoning puzzles designed to resist memorisation Novel abstraction and adaptation
SWE-Bench Verified Repository issues with human-validated test conditions Software repair
SWE-Bench Pro Harder and more diverse repository tasks Agentic coding
Terminal-Bench Tasks executed in a terminal environment Tool use, coding, systems work
BrowseComp Difficult information retrieval and browsing Search planning and synthesis
OSWorld Realistic computer interaction Desktop computer use
WebArena Tasks on realistic websites Web navigation and action
Vibe Code Bench End-to-end application construction Building a working product rather than a snippet
MMMU-Pro Multidisciplinary multimodal questions Visual and textual reasoning
FACTS Suite Grounding, parametric, search, and multimodal factuality Factual reliability
SimpleQA Verified Short factual questions with verified answers Parametric factual accuracy
HalluHard Difficult multi-turn conversations with citation checking Multi-turn hallucination under hard domains
MRCR Multi-round coreference and long-context retrieval Long-context recall and tracking
LongCodeBench Real code tasks at very long context Effective long-context coding
LongProc Long procedural generation with structured output Instruction retention and long-form coherence
HELM Holistic evaluation framework Multi-metric model comparison
MedHELM Clinical and medical tasks organised by workflow Health-domain applicability

Reading these names

Several entries differ from similarly named benchmarks in ways that matter. ARC-AGI is not the AI2 Reasoning Challenge. SWE-Bench Verified and SWE-Bench Pro are distinct constructions from the original SWE-Bench, and a score against one is not a score against another. A benchmark version must be recorded with every result, because these constructions are revised.

Full documentation for each, under the nine required headings, is in the benchmarks section. An entry appears in that section only after its construction methodology has been located and registered.