CESBench - Judgment: leaderboard
The 67 judgment items of CESBench (Zhou et al., 2026): the model labels a technical security claim true or false and justifies it. A wrong verdict scores zero and a right one earns the mean of four expert rubric points judged by Qwen3.5-397B, reported in percent; higher is better.
Metric: Judgment score (%; the 67 true-or-false security claims: a wrong verdict scores 0, a right one earns the mean of four justification rubric points judged 0, 0.5 or 1; 380 expert-written items on cryptographic engineering security for IoT devices; zero-shot, temperature 0, no tools, provider default reasoning mode; judgment and scenario answers scored by a Qwen3.5-397B judge from outside the evaluated set). Source: arxiv.org. Saturation forecast: Around March 2028. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.2 | 58.8 | #137 |
| 2 | Kimi K2.6 | 58 | #173 |
| 3 | Gemini 3.7 Flash | 54.9 | #32 |
| 4 | GPT-5.6 Luna | 54.5 | #117 |
| 5 | MiniMax-M2.5 | 48.3 | #590 |
| 6 | DeepSeek V4 Pro | 47 | #181 |
| 7 | Hunyuan A13B-Instruct | 43.3 | #2021 |
| 8 | DeepSeek V3 | 41.2 | #705 |
| 9 | Ling-flash-2.0 | 40.1 | #1352 |
| 10 | GLM-4 32B (0414) | 38.6 | #1407 |
| 11 | Llama 4 Maverick | 34.9 | #1105 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=cesbench-judgment · How It Works · Data refreshed daily, snapshot 2026-11-12.