finding
active
finding:the-between-to-within-class-variance-ratio-peaks-at-different-layers-for-different-tasks-confirming-no-single-layer-is-universally-optimalThe between-to-within-class variance ratio peaks at different layers for different tasks, confirming no single layer is universally optimal.
Supports the claim against single-layer probing approaches used in prior work.
Source paper
extracted_from(2026) · Angelos Poulis · Mark Crovella · Evimaria Terzi
Neighborhood — ranked by edge-count
Claims (1)
claim
- Argues against the single-layer analysis approach of prior work.
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.
- Undermines concept-based task definition in circuit discovery.
- Preference steering effective only in mid-layer range though correlations hold across layersfinding0.762Suggests mid-layer reading circuitry.
- At S=200 variation across runs (TVD 0.0451) nearly equals sampling variation within a run (TVD 0.0442).finding0.754Supports multinomial model.
- §5.5.
- Sparsity trade-off.
- A random head in the same layer shows almost no structure between count and width probesfinding0.748Control for twisting.
- Assumption D.3 formalized in the theoretical framework; empirically validated with coin-flip sequence experiments
- Layer selection.