finding
active
finding:j-lens-equipped-auditing-agent-performs-comparably-to-nla-equipped-on-four-of-six-cases-but-lags-on-reward-hacking-4-5-vs-10-0-and-language-switching-6-0-vs-9-9J-lens-equipped auditing agent performs comparably to NLA-equipped on four of six cases but lags on reward hacking (4.5 vs 10.0) and language switching (6.0 vs 9.9).
Automated auditing.
Source paper
extracted_fromAuthors
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Related work.
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.
- Demonstrates practical utility: NLAs enable root-cause discovery without access to misaligned model's training data.
- Methodological comparison.
- Tuned lens skips to output answers and is best next-token predictor while J-lens has highest KL.finding0.807Comparison.
- J-lens eval-awareness score separates NLA-flagged from unflagged positions with AUC 0.853.finding0.798Eval awareness probe.
- Model organism results.
- No-instruction baseline rate of target in J-lens is approximately zero across task families.finding0.775Modulation control.
- Downstream task validating NLA utility for model auditing; agents succeed without access to misalignment training data.
- Imagine instruction raises property name in J-lens while J-orthogonalized property probe stays at baseline.finding0.768J-space privileged.