IndustryBench - Selection and Substitution: leaderboard
Metric: Final (SV) score (out of 3): mean 0-3 rubric score from a Qwen3-Max judge against the reference answer, set to 0 for any answer with a safety violation against the source standard; closed-book, zero-shot, provider-default decoding, over the 649 items of the 2,049 Chinese industrial procurement QA items of IndustryBench grounded in Chinese national standards (GB/T) in the Selection and Substitution capability dimension (model selection, substitution and performance comparison; default non-reasoning runs); higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 17 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5.4 (Non-reasoning) | 2.19 |
| 2 | Claude Opus 4.6 | 2.12 |
| 3 | Gemini 3.1 Pro (Preview) | 2.11 |
| 4 | Qwen 3.5 Plus (Non-reasoning) | 2.08 |
| 5 | Qwen 3.5 397B A17B (Non-reasoning) | 2.08 |
| 6 | Qwen 3.6 Plus (Non-reasoning) | 2.07 |
| 7 | GPT-5.2 (Non-reasoning) | 2.03 |
| 8 | Qwen 3.5 122B A10B (Non-reasoning) | 2.03 |
| 9 | Kimi K2.5 (Non-reasoning) | 2 |
| 10 | Qwen 3.5 27B (Non-reasoning) | 1.92 |
| 11 | Claude Sonnet 4.6 | 1.89 |
| 12 | MiniMax-M2.5 | 1.85 |
| 13 | Qwen 3.5 35B A3B (Non-reasoning) | 1.8 |
| 14 | GLM-5 (Non-reasoning) | 1.78 |
| 15 | Qwen 3 235B A22B (Non-reasoning) | 1.55 |
Interactive version: theaggregate.ai/benchmark?slug=industrybench-selection-and-substitution · How It Works · Data refreshed daily, snapshot 2026-10-07.