GAIA
huggingface.co- Category
- Developer Tools
- Rank
- No. 3111Tools index
Previous survey · No. 2885 ·
- Listed in
- #28 Find AI benchmarks
- Pricing
- Free
- Type
- TOOL
- Use case
- Model & Agent Evaluation
- Date
About
A benchmark for general AI assistants from researchers at Meta and Hugging Face: 466 real-world questions that need reasoning, multi-modality, web browsing and tool use together, with 300 answers held back for the leaderboard and 166 released publicly. The gap it documented is the point — humans scored 92% where GPT-4 with plugins managed 15% — and it argues that robustness on questions humans find easy is a better milestone than ever-harder specialist tasks.
Why it made the leaderboard
466 questions that cannot be answered without chaining reasoning, browsing and tool use — the shape of real assistant work rather than a specialist exam. Its founding result frames the whole agent field: humans scored 92% where GPT-4 with plugins managed 15%, on questions people find straightforward.
Tags
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.