concept
active
concept:three-stage-layer-trajectoryThree-Stage Layer Trajectory
Empirically observed pattern in E3: early enrichment (ρd dips), mid-layer alignment (dr falls), late standardization (re-clustering)
Neighborhood — ranked by edge-count
Findings (3)
finding
- E3 result establishing the Goldilocks zone at mid-layers for LLaMA architecture
- E3 backbone generalization finding for Gemma; validates pattern across diverse architectures
- E3 backbone-specific finding showing three-stage trajectory generalizes across architectures
Concepts (1)
concept
- Goldilocks zoneassociated_withMiddle layers (e.g., 6-15) where anchoring is maximal; coined from metaphor.
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.
- Computing per-layer S(ℓ) to summarize geometry.
- Quantitative study correlating layer-wise anchoring geometry (S_max, AUS_N) with behavioral thresholds θ50
- Prior finding by Yang & Buzsaki and Campbell et al. on how deception representations evolve across layers; partially replicated and contrasted by this paper
- Plot of per-layer anchoring score S(ℓ) across model depth, revealing early dip, mid-layer peak, late standardization.
- Strategic filtering procedure that removes invalid trajectories and maintains optimal positive-to-negative trajectory ratio to stabilize training.
- Layer-wise trajectories show early enrichment, mid-layer alignment, and late re-clustering.claim0.745Qualitative geometry pattern.
- Procedure of systematically varying the layer at which activations are recorded and injected.
- First stage of DiffLogic CA update where each cell gathers information from neighboring cells via logic gate kernels