# BIG-Bench Hard

BIG-Bench Hard is an AI benchmark; this page records what it measures and traces its data, grader, and scores to primary sources.

Tests multi-step reasoning on 23 BIG-bench tasks where early language models trailed average human raters.

## Details

- Address: Online, Global
- Website: [github.com/suzgunmirac/BIG-Bench-Hard](<https://github.com/suzgunmirac/BIG-Bench-Hard>)
- Tags: Helpful, Science
- Category: Reasoning
- Status: Active
- Data licence: MIT
- Data access: open
- Grader: exact
- Organization: Suzgun et al.
- Released: 2022-10
- Paper: [arxiv.org/abs/2210.09261](<https://arxiv.org/abs/2210.09261>)
- Repository: [github.com/suzgunmirac/BIG-Bench-Hard](<https://github.com/suzgunmirac/BIG-Bench-Hard>)
- Task count: 23
- Top-score provenance: unknown
- ABI family: No

## Questions

### Where is BIG-Bench Hard?

BIG-Bench Hard is at Online, Global.

### How do I contact BIG-Bench Hard?

To contact BIG-Bench Hard, visit https://github.com/suzgunmirac/BIG-Bench-Hard.

## Source

From [Benchmark primary sources](https://aibenchmarkindex.com/about), license CC0-1.0, checked 2026-10-15.

Last updated October 15, 2026.

## Links

- This benchmark on the [web page](<https://aibenchmarkindex.com/global/online/big-bench-hard>)
- [Reasoning in Online](<https://aibenchmarkindex.com/global/online/category/reasoning.md>)
- [More benchmarks in Online](<https://aibenchmarkindex.com/global/online.md>)

> **Check before you compare.** Benchmark methods and leaderboards change. Follow the cited primary source before using a result in a decision.