finding
active
finding:orange-is-the-top-lens-token-while-sonnet-4-5-copies-a-sentence-under-the-instruction-to-concentrate-on-citrus-fruitsOrange is the top lens token while Sonnet 4.5 copies a sentence under the instruction to concentrate on citrus fruits.
Directed modulation.
Source paper
extracted_fromAuthors
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- Directed modulation prediction.
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.
- Introspection.
- Tuned lens skips to output answers and is best next-token predictor while J-lens has highest KL.finding0.757Comparison.
- Verbal report causal example.
- Imagine instruction raises property name in J-lens while J-orthogonalized property probe stays at baseline.finding0.742J-space privileged.
- Hidden computation.
- In blackmail transcript J-lens surfaces leverage, threat, shutdown, murder before any output token.finding0.741Silent strategic thinking.
- Pull-in on demand.
- Validates Assumption D.3 that instruction-tuned models prefer representative distributions, supporting the VS theoretical framework