Detecting Unintended Outputs via Introspection

{
  "doc": "lindsey-introspective-awareness-2026.md",
  "author": null,
  "context": "Models can distinguish artificially prefilled outputs from intentional responses by referencing prior internal representations; injection of matching concept vector causes model to retroactively accept prefill as intentional.",
  "enrichment": {
    "is_stale": true
  },
  "norm_label": "Detecting Unintended Outputs via Introspection",
  "source_url": null,
  "graphify_id": "prefill_detection_exp",
  "source_file": "lindsey-introspective-awareness-2026.md",
  "imported_from": "/Users/antonborzov/Documents/Research.nosync/papers/extract_typed_out/lindsey-introspective-awareness-2026/graph.json",
  "extracted_type": "finding",
  "source_location": "§7",
  "graphify_file_type": "finding"
}