finding
active
finding:steering-effect-between-concepts-correlates-with-my-centroid-distance-r-65-p-001-for-emotionsSteering effect between concepts correlates with My centroid distance, r=.65 (p<.001) for Emotions
Geometry predicts steering.
Source paper
extracted_from(2026) · Eric Bigelow · Raphaël Sarfati · Daniel Wurgaft · Owen Lewis +4
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- Steering effects will be structured by manifold geometryassociated_withsupportsPredictive claim tested in Fig 6.
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.
- Weaker geometry-steering correlation.
- Emotion centroid distances in My correlate r=.93 with human valence-arousal ratings (Russell & Mehrabian)finding0.860LLM emotion geometry matches humans.
- Steering effect size across 35 emotion vectors correlates with probe-preference correlation at r=0.85finding0.831Emotions correlating with preference causally drive it.
- Steering between concept centroids on Llama3.1-8B smoothly and predictably changes token probabilitiesfinding0.808Shows manifolds causally influence behavior.
- Second significant cross-concept introspection improvement; marginal after BH correction (q≈0.066)
- Strongest cross-concept introspection improvement; survives BH correction (q≈0.011)
- Off-target steering effect vanishes for concepts outside the top-level cluster in Emotions and Genresfinding0.780Hierarchy explains entanglement.
- Finding that the two evaluation modalities frequently diverge in their interpretation of the same SAE feature