finding
active
finding:swapping-probe-j-space-components-flips-answer-on-61-of-90-two-hop-prompts-versus-28-for-non-j-space-which-falls-to-6-with-clampSwapping probe J-space components flips answer on 61% of 90 two-hop prompts versus 28% for non-J-space, which falls to 6% with clamp.
Privilege in reasoning.
Source paper
extracted_fromAuthors
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Probe decomposition.
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.
- Privileged report.
- Mediation control.
- Swapping pure J-lens vectors drives swap target into top-5 on 88% of trials on Sonnet 4.5.finding0.778Scaled causal report.
- Mechanistic proof of concept.
- Hypothesis generation.
- Shows the passive vs. active divide is more important than the specific wording of instructions.
- Reward-driven decision.
- J-space component carries a median of only 6-7% of concept vector variance on Sonnet 4.5.finding0.758Small variance share.