finding
active
finding:steering-between-concept-centroids-on-llama3-1-8b-smoothly-and-predictably-changes-token-probabilitiesSteering between concept centroids on Llama3.1-8B smoothly and predictably changes token probabilities
Shows manifolds causally influence behavior.
Source paper
extracted_from(2026) · Usha Bhalla · Thomas Fel · Can Rager · Sheridan Feucht +8
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Manifold structure is present in representations and causally influences downstream behaviorsupportsInterpretation of steering results.
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.
- Core empirical result demonstrating that manifold steering produces on-target, behavior-aligned outputs.
- Variance decomposition showing AS results are dominated by persona identity
- Empirical result demonstrating the failure mode of linear steering when concept geometry is cyclic.
- Second significant cross-concept introspection improvement; marginal after BH correction (q≈0.066)
- Steering effect between concepts correlates with My centroid distance, r=.65 (p<.001) for Emotionsfinding0.808Geometry predicts steering.
- Strongest cross-concept introspection improvement; survives BH correction (q≈0.011)
- Illustrative finding that ESR mitigates but does not fully eliminate steering influence
- Optimal steering layers: Llama-3.1-8B layer 12, Gemma-2-9B layer 20, Qwen-2.5-7B layer 14finding0.788Layer sweep result.