泄露的 AI 基准测试报告照片

{ "type": "显示学术技术报告的电脑显示器照片", "style": "轻微倾斜的屏幕照片，可见摩尔纹，LCD 像素网格，轻微反光，LaTeX 文档排版，衬线字体", "document_header": { "left": "4 基准评估", "right": "DeepSeek-V4 技术报告" }, "introductory_text": "总结 DeepSeek-V4 对比 GPT-5.3、Claude Opus 4.6 以及 Gemini 3.1 Pro Preview 的综合评估段落。", "visualizations": { "legend": "5 个带有颜色代码的项目：深蓝色、灰色、浅灰色、蓝色条纹、浅蓝色", "bar_charts": { "count": 6, "labels": [ "MMLU-Pro (EM)", "GPQA-Diamond (Pass@1)", "AIME 2025 (Pass@1)", "LiveCodeBench (Pass@1-COT)", "SWE-bench Verified (Resolved)", "Tau-bench (Average)" ] }, "caption": "图 1 | 核心基准测试性能对比。DeepSeek-V4 在大多数基准测试中均取得了业界领先的成果。" }, "data_table": { "columns": [ "基准测试", "DeepSeek-V4", "GPT-5.3", "Claude Opus 4.6", "Gemini 3.1 Pro Preview", "GPT-4.1" ], "categories": { "count": 4, "rows": [ {"label": "通用", "icon": "globe/network", "sub_items": 3}, {"label": "推理与数学", "icon": "calculator/clipboard", "sub_items": 3}, {"label": "代码", "icon": "code brackets", "sub_items": 3}, {"label": "Agent", "icon": "robot face", "sub_items": 3} ] } } }