Automated auditing benchmark requiring end-to-end investigation of intentionally-misaligned model; NLA-equipped agents outperform baselines.

{
  "doc": "natural.md",
  "context": "Downstream task validating NLA utility for model auditing; agents succeed without access to misalignment training data.",
  "category": "ai",
  "norm_label": "Automated auditing benchmark requiring end-to-end investigation of intentionally-misaligned model; NLA-equipped agents outperform baselines.",
  "graphify_id": "evaluation_benchmark",
  "source_file": "natural.md",
  "imported_from": "/Users/antonborzov/Documents/Research.nosync/papers/extract_typed_out/natural/graph.json",
  "extracted_type": "finding",
  "source_location": "§Introduction, §Evaluating NLAs during training",
  "graphify_file_type": "finding"
}