WorldCoder-Bench - Robust: leaderboard
Verification coverage on WorldCoder-Robust, the perturbation split of WorldCoder-Bench (2026): each core task is re-instantiated three times with randomized physical constants, object counts, initial states, prompt phrasing and asset names and a rewritten hidden contract, so recalled templates fail. Percent; higher is better.
Metric: Verification coverage (%) on WorldCoder-Robust, 615 perturbed variants of the WorldCoder-Core tasks (three per task, with randomized physical constants, object counts, initial states, prompt phrasing and asset filenames and rewritten contracts), same zero-shot protocol and StateProbe checks; higher is better. Source: arxiv.org. Saturation forecast: Around July 2028. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 19.9 | #123 |
| 2 | Gemini 3.1 Pro (Preview) | 19.1 | #94 |
| 3 | Kimi K2.5 | 15.6 | #256 |
| 4 | Claude Opus 4.6 | 13.9 | #102 |
| 5 | Qwen 3.6 Plus | 13.5 | #222 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=worldcoder-bench-robust · How It Works · Data refreshed daily, snapshot 2026-11-13.