finding
active
finding:probability-of-same-answer-under-paraphrasing-shows-a-smooth-upward-curve-for-most-tasks-not-flat-or-downward-as-steganography-predicts-aqua-and-logiqa-slight-exceptions-on-six-step-samplesProbability of same answer under paraphrasing shows a smooth upward curve for most tasks, not flat or downward as steganography predicts; AQuA and LogiQA slight exceptions on six-step samples
Appendix B.
Source paper
extracted_from(2023) · Tamera Lanham · Anna Chen · Ansh Radhakrishnan · Benoit Steiner +26
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- From paraphrasing results.
Hypotheses (1)
hypothesis
- If steganography were important, paraphrasing midway would corrupt the encoded message, giving flat or downward-sloping lines.associated_withAppendix B 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.
- Retrodiction of known phenomenon.
- Little evidence of steganography in NLAs; meaning-preserving transformations cause only small drops in FVEfinding0.780Quantitative evaluation showing NLAs do not heavily rely on covert encoding beyond overt language.
- Key improvement in cross-task generalization enabled by explicit instruction framing.
- Smoothed behavior.
- Shows that explicit instructions delay the emergence of truth directions in arithmetic tasks.
- Table 2, row 3, showing equivalence when prior preferences match rewards.
- Fig. 7d,e.
- Explains discrepancy.