Layer 18

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• early layers (0–5)concept0.735
  Layers with weak anchoring due to generic representations.
• Layer Normalization (Ba et al., 2016b)concept0.712
  Layer normalisation used in transformer and in TEM-t position encoding preprocessing.
• Math/code tasks S ≈ -1.65 at layers 8–12finding0.709
  Task-specific peak anchoring score for structured reasoning domains.
• deeper layers (16–28)concept0.704
  Layers where anchoring weakens systematically due to representational drift.
• Layer-Incremental Vectorconcept0.699
  A formulation for multi-layer steering where each layer's contribution is the incremental difference in persona direction representation
• Layer sweepmethod0.693
  Procedure of systematically varying the layer at which activations are recorded and injected.
• Q8B most often peaks at layer 20, followed by layer 25; G20B most often peaks at layer 15finding0.684
  Best steering layer is model-specific; linearly accessible trait information is organized differently across models
• Multi Layer Perceptronframework0.681
  Network with hidden layers capable of representing non-linearly separable functions, enabling deep model induction