finding
active
finding:j-lens-outperforms-logit-and-tuned-lens-on-pass-k-auc-across-six-prompt-distributions-and-has-about-twice-the-ablation-kl-on-multihopJ-lens outperforms logit and tuned lens on pass@k AUC across six prompt distributions and has about twice the ablation KL on multihop.
Methodological comparison.
Source paper
extracted_fromAuthors
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Methods comparison.
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.
- Tuned lens skips to output answers and is best next-token predictor while J-lens has highest KL.finding0.860Comparison.
- Automated auditing.
- Model organism results.
- Related work.
- Mechanistic proof of concept.
- J-lens eval-awareness score separates NLA-flagged from unflagged positions with AUC 0.853.finding0.766Eval awareness probe.
- Logit lens prediction accuracy is near-chance at layer 4 (28%) after injection at L2, α=6finding0.762Shows that signal integration into explicit prediction has barely begun immediately after injection
- No-instruction baseline rate of target in J-lens is approximately zero across task families.finding0.758Modulation control.