Rapport Technique de Benchmark sur Écran

{
  "type": "photograph of a computer monitor displaying an academic technical report",
  "style": "slightly angled screen photo, visible moire pattern, LCD pixel grid, subtle screen glare, LaTeX document formatting, serif typography",
  "document_header": {
    "left": "4 Benchmark Evaluation",
    "right": "[DeepSeek-V4 Technical Report]"
  },
  "introductory_text": "Paragraph summarizing comprehensive evaluation of [DeepSeek-V4] against [GPT-5.3], [Claude Opus 4.6], and [Gemini 3.1 Pro Preview].",
  "visualizations": {
    "legend": "5 items with color codes: dark blue, grey, light grey, blue striped, light blue",
    "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": "Figure 1 | Performance comparison on core benchmarks. [DeepSeek-V4] achieves state-of-the-art results across the majority of benchmarks."
  },
  "data_table": {
    "columns": [
      "Benchmark",
      "[DeepSeek-V4]",
      "[GPT-5.3]",
      "[Claude Opus 4.6]",
      "[Gemini 3.1 Pro Preview]",
      "GPT-4.1"
    ],
    "categories": {
      "count": 4,
      "rows": [
        {
          "label": "General",
          "icon": "globe/network",
          "sub_items": 3
        },
        {
          "label": "Reasoning & Math",
          "icon": "calculator/clipboard",
          "sub_items": 3
        },
        {
          "label": "Code",
          "icon": "code brackets",
          "sub_items": 3
        },
        {
          "label": "Agent",
          "icon": "robot face",
          "sub_items": 3
        }
      ]
    }
  }
}