SteelBench - Compositional Reasoning Gap: leaderboard
Metric: Compositional reasoning gap (0-1; share of worker instances with a correctly recognised action whose safety-violation judgment is still wrong, lower is better; 1,345 densely annotated clips of real steel-plant CCTV footage (8 frames at 1080p per clip) from six industrial domains, zero-shot structured prompt with the 25-class action taxonomy, site context, PPE definitions and safety rules; provider-default temperature; failed parses count as wrong). Source: arxiv.org. Saturation forecast: Around 2031. 9 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Llama 4 Maverick | 0.38 |
| 2 | GPT-4o | 0.39 |
| 3 | GPT-5.4 | 0.48 |
| 4 | Gemini 2.5 Pro | 0.5 |
| 5 | Gemini 2.5 Flash | 0.5 |
| 6 | Nemotron Nano 12B v2 VL | 0.52 |
| 7 | Gemma 4 31B (IT) | 0.55 |
| 8 | Claude Opus 4.7 | 0.57 |
| 9 | Qwen 3.5 122B A10B | 0.58 |
Interactive version: theaggregate.ai/benchmark?slug=steelbench-compositional-reasoning-gap · How It Works · Data refreshed daily, snapshot 2026-09-29.