{
  "cell_id": "q115_hard_negative_k2",
  "doc_id": "2023.findings-emnlp.248.pdf",
  "question": "In the paper main results table comparing Standard, Chain-of-Thought, and the proposed method across the five logical reasoning datasets, on which dataset does the proposed method (without self-refinement) achieve its highest accuracy when GPT-4 is the underlying LLM?",
  "answer": "LogicalDeduction",
  "answer_format": "Str",
  "condition": "hard_negative",
  "k": 2,
  "source_position": "first",
  "modality": "image",
  "dpi": 150,
  "image_docs": [
    {
      "role": "source",
      "doc": "2023.findings-emnlp.248",
      "pages": 19,
      "image_dir": "images/2023.findings-emnlp.248"
    },
    {
      "role": "hn_1",
      "doc": "web_39b6cadc4e9ca03e",
      "pages": 27,
      "image_dir": "images/web_39b6cadc4e9ca03e"
    },
    {
      "role": "hn_2",
      "doc": "web_949ca008fd6aab55",
      "pages": 12,
      "image_dir": "images/web_949ca008fd6aab55"
    }
  ],
  "n_images": 58
}