finding
active
finding:hostile-probe-activation-anti-correlates-with-preference-at-r-0-74Hostile probe activation anti-correlates with preference at r=-0.74
Negative probe predicts dispreference.
Source paper
extracted_from(2026) · Nicholas Sofroniew · Isaac Kauvar · William Saunders · Runjin Chen +12
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Key finding.
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.
- Positive valence probe predicts preference.
- Discovery claim.
- Assistant-colon probe values predict response emotion better than user-turn (r=0.87 vs 0.59)finding0.763Colon token carries prepared emotion.
- User and Assistant emotions dissociate.
- Primary quantitative result: probe method outperforms gradient-based and LLM-judge alternatives at lower computational cost.
- Valence mediates preference.
- Internals-based approach to which the logit monitor is complementary.
- Justifies restricting probe-based vector derivation to h_b activations; attributed to Yes/No semantics