finding
active
finding:logit-lens-decodes-probe-directions-to-cheating-illicit-dishonest-tokens-though-absent-from-synthetic-dataLogit lens decodes probe directions to cheating/illicit/dishonest tokens though absent from synthetic data
Probes encode cheating concept.
Source paper
extracted_from(2026) · Leon Bergen · Usha Bhalla · Andrew Lee · Barak Widawsky +14
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Interpretation of logit lens, exemplars, steering.
Questions (1)
question
- Open scientific question motivating the 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.
- Central methodological contribution: computing probability-weighted expected value over digit-token logits recovers continuous, informative signal
- Pull-in on demand.
- Logit lens prediction accuracy is near-chance at layer 4 (28%) after injection at L2, α=6finding0.745Shows that signal integration into explicit prediction has barely begun immediately after injection
- Mechanistic.
- In blackmail transcript J-lens surfaces leverage, threat, shutdown, murder before any output token.finding0.738Silent strategic thinking.
- Tuned lens skips to output answers and is best next-token predictor while J-lens has highest KL.finding0.738Comparison.
- Model organism results.
- Model organism detection.