Five prediction tasks improve with NLA training across three models (Opus 4.6, Haiku 4.5, Haiku 3.5).

{
  "doc": "natural.md",
  "context": "Systematic evidence that NLA explanations become more informative during training despite optimizing only for reconstruction.",
  "category": "ai",
  "norm_label": "Five prediction tasks improve with NLA training across three models (Opus 4.6, Haiku 4.5, Haiku 3.5).",
  "graphify_id": "quantitative_evals_5_tasks",
  "source_file": "natural.md",
  "imported_from": "/Users/antonborzov/Documents/Research.nosync/papers/extract_typed_out/natural/graph.json",
  "extracted_type": "finding",
  "source_location": "§Evaluating NLAs during training",
  "graphify_file_type": "finding"
}