finding
active
finding:steering-at-6-layers-strength-0-6-each-total-3-6-outperforms-single-layer-steering-at-equivalent-total-strength-for-type-hint-suppressionSteering at 6 layers (strength 0.6 each, total 3.6) outperforms single-layer steering at equivalent total strength for type hint suppression
Demonstrates distributed steering is more effective and less accuracy-damaging than concentrated steering.
Source paper
extracted_from(2025) · Hua, Tim Tian · Qin, Andrew · Marks, Samuel · Nanda, Neel
Neighborhood — ranked by edge-count
Claims (1)
claim
- Practical finding for optimizing steering setup.
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.
- Self-repair effect.
- Suggests localization.
- Four best contrastive prompt pairs outperform full 16-pair average steering vector for type hint suppressionfinding0.822Optimization result for steering vector construction.
- Demonstrates averaging multiple prompt pairs reduces noise; optimal subset selection further improves performance.
- Contrast with probe steering.
- Preference steering effective only in mid-layer range though correlations hold across layersfinding0.804Suggests mid-layer reading circuitry.
- Extends single-layer results to show multi-layer steering is more effective for difficult cases
- Optimal steering layers: Llama-3.1-8B layer 12, Gemma-2-9B layer 20, Qwen-2.5-7B layer 14finding0.785Layer sweep result.