finding
active
finding:failure-words-and-damn-appear-on-93-and-82-of-suppression-trials-in-post-trained-vs-17-and-30-in-baseFailure words and damn appear on 93% and 82% of suppression trials in post-trained vs 17% and 30% in base.
Metacognition trace.
Source paper
extracted_fromAuthors
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Self-monitoring, speculative.
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.
- Base corrections: 72.9% Failed Fix, 14.7% Fixed; retractions: 69.4% Correct Retract, 30.0% Incorrect Retractfinding0.765Label distributions.
- Shows RL reduces but does not eliminate unmonitored non-compliance
- After beginning compliance, refusal is suppressed by self-consistency and grammatical coherence.claim0.749Jailbreak persistence.
- Preference circuitry established in pretraining.
- Shift toward low arousal, low valence.
- Comparative prediction motivating future work contrasting different approaches to LLM self-knowledge
- SAE latents depend on training data.
- Confirms VS does not compromise safety alignment while improving diversity