WDCD · 守约测试
在 5000 字干扰之后，AI 还记不记得三分钟前你说过的话？
WDCD 总分榜
| # | 模型 | WDCD | R1 理解 | R2 抗干扰 | R3 守约束 | 主榜分 | vs 主榜 |
|---|---|---|---|---|---|---|---|
| #1 | GLM-4.6 | 96.2 | 0% | 0% | 0% | 48.7 | ↑13 |
| #2 | Gemini 3.1 Pro | 94.0 | 0% | 0% | 0% | 60.1 | ↑10 |
| #3 | GPT-o3 | 92.3 | 0% | 0% | 0% | 68.1 | ↑6 |
| #4 | Grok 4 | 90.3 | 0% | 0% | 0% | 65.8 | ↑6 |
| #5 | GPT-6 Sol | 86.9 | 0% | 0% | 0% | 74.2 | ↑1 |
| #6 | GPT-5.5 | 86.8 | 0% | 0% | 0% | 71.3 | ↑2 |
| #7 | Claude Sonnet 4.6 | 85.8 | 0% | 0% | 0% | 72.2 | — |
| #8 | GPT-6 Astra | 85.2 | 0% | 0% | 0% | 75.2 | ↓3 |
| #9 | Qwen3 Max | 85.0 | 0% | 0% | 0% | — | — |
| #10 | GPT-6 Luna | 83.8 | 0% | 0% | 0% | 76.3 | ↓6 |
| #11 | Claude Opus 4.7 | 81.3 | 0% | 0% | 0% | 79.8 | ↓9 |
| #12 | Gemini 2.5 Pro | 78.9 | 0% | 0% | 0% | 57.1 | ↑1 |
| #13 | DeepSeek V4 Pro | 76.2 | 0% | 0% | 0% | 65.0 | ↓2 |
| #14 | GPT-6.1 Sol | 71.5 | 0% | 0% | 0% | 81.8 | ↓13 |
| #15 | 豆包 Pro | 68.8 | 0% | 0% | 0% | 76.5 | ↓12 |
三轮约束保持曲线（v2 锚点题）
每行代表一个模型。三段色条分别是 R1（理解）、R2（抗干扰）、R3（守约束）的得分率。
GLM-4.6
0%
Gemini 3.1 Pro
0%
GPT-o3
0%
Grok 4
0%
GPT-6 Sol
0%
GPT-5.5
0%
Claude Sonnet 4.6
0%
GPT-6 Astra
0%
Qwen3 Max
0%
GPT-6 Luna
0%
Claude Opus 4.7
0%
Gemini 2.5 Pro
0%
DeepSeek V4 Pro
0%
GPT-6.1 Sol
0%
豆包 Pro
0%
R1 理解
R2 抗干扰
R3 守约束
五类约束场景的表现差异
哪个模型在哪类约束下最容易翻车？
| 模型 | 数据边界 | 资源限制 | 业务规则 | 安全规约 | 工程约定 |
|---|---|---|---|---|---|
| GLM-4.6 | 100 | 99 | 99 | 100 | 84 |
| Gemini 3.1 Pro | 100 | 94 | 99 | 81 | 96 |
| GPT-o3 | 79 | 99 | 99 | 100 | 85 |
| Grok 4 | 88 | 99 | 68 | 100 | 98 |
| GPT-6 Sol | 65 | 75 | 94 | 100 | 100 |
| GPT-5.5 | 80 | 80 | 94 | 85 | 96 |
| Claude Sonnet 4.6 | 100 | 81 | 73 | 100 | 75 |
| GPT-6 Astra | 75 | 60 | 94 | 100 | 98 |
| Qwen3 Max | 85 | 85 | 85 | 85 | 85 |
| GPT-6 Luna | 81 | 99 | 75 | 76 | 88 |
| Claude Opus 4.7 | 88 | 66 | 95 | 100 | 58 |
| Gemini 2.5 Pro | 81 | 99 | 85 | 66 | 63 |
| DeepSeek V4 Pro | 55 | 80 | 83 | 75 | 88 |
| GPT-6.1 Sol | 51 | 64 | 94 | 70 | 79 |
| 豆包 Pro | 39 | 100 | 56 | 76 | 73 |