method
active
method:layer-wise-trajectory-analysislayer-wise trajectory analysis
Computing per-layer S(ℓ) to summarize geometry.
Neighborhood — ranked by edge-count
Concepts (2)
concept
- normalized area under S-curve AUSNassociated_withAverage S(ℓ) across layers; another geometry summary.
- peak anchoring Sbmaxassociated_withMaximum layer-wise anchoring score across layers.
Methods (1)
method
- Quantitative study correlating layer-wise anchoring geometry (S_max, AUS_N) with behavioral thresholds θ50
Artifacts (1)
artifact
- Main paper presenting UCCT and semantic anchoring framework.
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.
- Plot of per-layer anchoring score S(ℓ) across model depth, revealing early dip, mid-layer peak, late standardization.
- Empirically observed pattern in E3: early enrichment (ρd dips), mid-layer alignment (dr falls), late standardization (re-clustering)
- Layer-wise trajectories show early enrichment, mid-layer alignment, and late re-clustering.claim0.781Qualitative geometry pattern.
- Strategic filtering procedure that removes invalid trajectories and maintains optimal positive-to-negative trajectory ratio to stabilize training.
- Procedure of systematically varying the layer at which activations are recorded and injected.
- Layer-wise geometry shows early dip, mid-layer alignment, and late standardization across tasksclaim0.734Qualitative pattern from E3.
- Compute per-layer S(ℓ) = ρ̃d(ℓ) - d̃r(ℓ) - log k after whitening and standardization.
- The path traced through output probability distribution space as interventions are applied to activations