finding
active
finding:optimal-steering-layers-llama-3-1-8b-layer-12-gemma-2-9b-layer-20-qwen-2-5-7b-layer-14Optimal steering layers: Llama-3.1-8B layer 12, Gemma-2-9B layer 20, Qwen-2.5-7B layer 14
Layer sweep result.
Source paper
extracted_from(2025) · Eric Bigelow · Daniel Wurgaft · YingQiao Wang · Noah D. Goodman +3
Neighborhood — ranked by edge-count
Papers (1)
paper
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.
- E3 result establishing the Goldilocks zone at mid-layers for LLaMA architecture
- Steering between concept centroids on Llama3.1-8B smoothly and predictably changes token probabilitiesfinding0.788Shows manifolds causally influence behavior.
- Math and code tasks show strongest mid-layer anchoring on LLaMA (S ≈ −1.65 at layers 8-12)finding0.788Task-specific E3 finding showing compositional reasoning requires deeper processing
- Demonstrates distributed steering is more effective and less accuracy-damaging than concentrated steering.
- Core empirical result demonstrating that manifold steering produces on-target, behavior-aligned outputs.
- Establishes generalizability of the core difficulty-boundary finding across model families.
- Connects this study's results to Schrimpf et al. 2021 and Caucheteux et al. 2022/2023 findings on brain-LLM alignment.
- Self-repair effect.