finding
active
finding:inhibiting-key-side-correct-answer-features-inhibits-b-injecting-them-on-another-option-flips-output-to-that-letterInhibiting key-side correct-answer features inhibits B; injecting them on another option flips output to that letter
Perturbation validation.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- Prior hypothesis now confirmed.
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.
- Steering validation for antonym mechanism.
- QK attribution for MC heads.
- Supports monosemantic reading.
- Tool-induced answer flips mostly toward 'no effect' (NLA 82% positive, 84% negative), cancelling in accuracyfinding0.780Direction of tool-driven prediction changes.
- QK attribution of antonym heads.
- Causal validation.
- Default refusal circuit.
- Suggestive evidence for language-independent truth representation in LLMs