finding
active
finding:sae-feature-with-j-lens-readout-fake-fraud-suspicious-detects-fabricated-content-suppressing-it-cuts-test-recognition-from-28-50-to-10-50SAE feature with J-lens readout fake, fraud, suspicious detects fabricated content; suppressing it cuts test-recognition from 28/50 to 10/50.
Mechanistic.
Source paper
extracted_fromAuthors
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Ablation surfaced misaligned propensity.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- J-lens at fabricated search results shows fake, injection, false and 假 while model silently omitted them.finding0.822Silent recognition.
- Model organism detection.
- Model organism results.
- Related work.
- Imagine instruction raises property name in J-lens while J-orthogonalized property probe stays at baseline.finding0.772J-space privileged.
- Tuned lens skips to output answers and is best next-token predictor while J-lens has highest KL.finding0.769Comparison.
- J-lens eval-awareness score separates NLA-flagged from unflagged positions with AUC 0.853.finding0.769Eval awareness probe.
- Onset marker.