From Crowdsourced Data to High-Quality Benchmarks: Arena-Hard and BenchBuilder Pipeline
Summary
BenchBuilder is an automated pipeline that uses language models to curate hard, open-ended prompts from large crowd-sourced sets, so a benchmark can be updated without a human in the loop. They apply BenchBuilder to datasets such as Chatbot Arena and WildChat-1M, extracting challenging prompts and utilizing LLM-as-a-Judge for automatic model evaluation. They release Arena-Hard-Auto, a benchmark consisting 500 challenging prompts curated by BenchBuilder. They report 3x higher separation of model performances than MT-Bench and 98.6% correlation with human preference rankings, at a stated cost of $20. Code is on GitHub. [S1]
Claims
- id: C1 claim: "BenchBuilder automatically curates hard, open-ended prompts from crowd-sourced datasets." source: S1 evidence: "curate high-quality, open-ended prompts from large, crowd-sourced datasets"
- id: C2 claim: "Arena-Hard-Auto has 500 prompts, 3x the separation of MT-Bench, and 98.6% correlation with human rankings." source: S1 evidence: "3x higher separation of model performances compared to MT-Bench"
- id: C3 claim: "The authors release the benchmark code on GitHub." source: S1 evidence: "https://github.com/lmarena/arena-hard-auto"
Verification
- Code available: Paper: https://github.com/lmarena/arena-hard-auto, HTTP 200 on 2026-11-12, redirecting to https://github.com/Arena-OSS/arena-hard-auto. Not run.
- Reproduced: not reproduced (2026-11-12). Code not run.
- Related work: Related eval notes include 2403-04132-chatbot-arena and 2306-05685-mt-bench.
- Limits: The 98.6% correlation and $20 cost figures are from the paper's v2 setup. LLM-as-a-judge can inherit judge-model biases. The prompt set is 500 items, not a full user distribution.
- License: CC BY 4.0 on the abs page (2026-11-12; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 17 Jun 2024 (v1); last revised v2 (14 Oct 2024). Sources: [S1] https://arxiv.org/abs/2406.11939, v2 (14 Oct 2024). checked_on: 2026-11-12.
When to apply
Use when asking how Arena-Hard was built from Chatbot Arena data, or how it compares with MT-Bench.
Facts
- id: F1 claim: "BenchBuilder curates challenging prompts from crowd-sourced data such as Chatbot Arena and WildChat-1M." confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
- id: F2 claim: "Arena-Hard-Auto contains 500 prompts and is reported to have 3x MT-Bench's model separation and 98.6% correlation with human rankings." confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
- id: F3 claim: "Code was pointed to https://github.com/lmarena/arena-hard-auto." confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
Sources
- [S1] arXiv 2406.11939, https://arxiv.org/abs/2406.11939. License: CC BY 4.0. v2 (14 Oct 2024). Accessed 2026-11-12.
License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.
License: CC-BY-4.0 · Summary by iinventedlight (AI-assisted). Source: Tianle Li, Wei-Lin Chiang, Evan Frick, and 5 coauthors, From Crowdsourced Data to High-Quality Benchmarks: Arena-Hard and BenchBuilder Pipeline (CC BY 4.0).
Cite this
'iinventedlight (AI-assisted)'. 2026. "From Crowdsourced Data to High-Quality Benchmarks: Arena-Hard and BenchBuilder Pipeline." iinventedlight Library, iil:note:qkx33n1c, version of 2026-11-12. https://iinventedlight.com/library/notes/2406-11939-arena-hard/. License: CC-BY-4.0.
@misc{iil_note_qkx33n1c,
author = {{'iinventedlight (AI-assisted)'}},
title = {{From Crowdsourced Data to High-Quality Benchmarks: Arena-Hard and BenchBuilder Pipeline}},
howpublished = {iinventedlight Library},
year = {2026},
date = {2026-11-12},
url = {https://iinventedlight.com/library/notes/2406-11939-arena-hard/},
note = {Permanent ID iil:note:qkx33n1c. License: CC-BY-4.0.},
}
Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.
For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/2406-11939-arena-hard.md.