Benchmarking the capability spectrum and reliability of System One models
- Verified labels
- Difficulty tiers
- Agent sandboxes
- Reliability tests
01 — At a glance
Capability, speed and reliability
02 — Leaderboard
Leaderboard
The mean of 11 domains, each on its hardest tier.
03 — Decision reliability
Does a decision hold when the options change?
SOption-order stability: keeps its answer when the options are shuffled
NNull discrimination: notices the missing answer
RReliability = √(S·N), from 0 to 100
Reliability ranking
04 — Domains
Where System One matches System Two, and what closes the gap
How big is the gap to System Two, and where is it largest?
05 — The benchmark
2,919 tasks in 11 domains
Static questions plus web and tool sandboxes, all posed as decisions.
How the leaderboard is scored
Domains count equally, each on its hardest tier: 1,232 tasks.
ci(m) ∈ [0, 1] is model m's credit on task i; Td holds domain d's leaderboard tasks; D = 11.
06 — Evaluate
Evaluate your model
Three steps, any model.
Step 01 · Get the data
git clone https://github.com/holdrain/SO-Arena
cd SO-Arena
./jev fetch-data --output /path/to/so-arena-data
export SO_ARENA_DATA=/path/to/so-arena-data
./jev verify
./jev setupDownloads and checks the data.
Step 02 · Run your model
export JEV_ENDPOINT=http://localhost:8000
./jev run --adapter adapters:http \
--model your-model --out /path/to/new-runPoint it at your model server.
Step 03 · Score
./jev score --run /path/to/new-run \
--output /path/to/report.jsonThe leaderboard score, by domain and tier.
Full instructions: the SO-Arena repository.
07 — Citation
Cite SO-Arena
Correspondence: xinchao@nus.edu.sg
@misc{li2026soarena,
title = {SO-Arena: Benchmarking the Capability Spectrum and Reliability of System One Models},
author = {Lichen Li and Ziping Dong and Sicheng Feng and Qi Li and Xinchao Wang},
year = {2026},
eprint = {2611.15777},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.15777}
}