finding
active
finding:arbitrary-domain-shows-lower-on-target-and-higher-off-target-steering-effects-and-my-does-not-predict-entanglementArbitrary domain shows lower on-target and higher off-target steering effects, and My does not predict entanglement
Counter-example.
Source paper
extracted_from(2026) · Eric Bigelow · Raphaël Sarfati · Daniel Wurgaft · Owen Lewis +4
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Interpretation of Arbitrary 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.
- Justification for the novel metric introduced in the paper
- Core empirical claim comparing steering approaches on cyclic concepts.
- Result categorizing concept steerability into three distinct regimes.
- Empirical result demonstrating the failure mode of linear steering when concept geometry is cyclic.
- Off-target steering effect vanishes for concepts outside the top-level cluster in Emotions and Genresfinding0.766Hierarchy explains entanglement.
- Empirical demonstration on Llama-3.1-8B that steering along representation manifold aligns outputs with behavior manifold, whereas linear steering does not.
- A linear model predicts out-of-sample steering effects on queried concepts from manifold distancesfinding0.760Predictive geometry.
- DIM steering has less off-target effect but smaller maximum on-target effect than linear probe steeringfinding0.759Method tradeoff.