claim
active
claim:random-word-prefix-prompts-show-emergence-patterns-similar-to-no-prompt-suggesting-prompt-length-alone-does-not-shift-truth-geometryRandom word prefix prompts show emergence patterns similar to no-prompt, suggesting prompt length alone does not shift truth geometry.
Control experiment ruling out token-count as the cause of truth geometry shifts.
Source paper
extracted_from(2026) · Angelos Poulis · Mark Crovella · Evimaria Terzi
Neighborhood — ranked by edge-count
Claims (1)
claim
- Supported by the neutral read-prompt changing emergence but not fully replicating ask-correct cross-task generalization.
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.
- Control prompt with random words of same length as ask-correct to isolate token-count confounds.
- From the cross-task generalization heatmaps in Appendix B.3.3.
- Refusal is not a single linear classifier.
- Suggestive evidence for language-independent truth representation in LLMs
- Supports the view that R1's boundaries differ from user-perceived ones.
- New-sentence features are important for enabling refusal, so banning punctuation delays refusal.hypothesis0.769Predicts no-punctuation instruction makes jailbreak more effective.
- QK attribution of antonym heads.
- A post-MLP feature for words indicating uniqueness has pre-MLP inputs firing on specific such wordsfinding0.762Abstract features built from specific instances.