claim
active
claim:single-layer-steering-fails-because-of-model-self-repair-at-later-layersSingle-layer steering fails because of model self-repair at later layers
Interpretation of Fig 13-14.
Source paper
extracted_from(2026) · Eric Bigelow · Raphaël Sarfati · Daniel Wurgaft · Owen Lewis +4
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (2)
finding
- Contrast with probe steering.
- Self-repair effect.
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.
- Suggests localization.
- Demonstrates distributed steering is more effective and less accuracy-damaging than concentrated steering.
- Practical finding for optimizing steering setup.
- Central interpretive claim organizing the entire paper's results
- First steering insight specific to R1.
- Preference steering effective only in mid-layer range though correlations hold across layersfinding0.743Suggests mid-layer reading circuitry.
- The paper's critique of the standard linear steering baseline, supported by the days-of-week demo.
- Explanation of brittle steering.