finding
active
finding:a-random-head-in-the-same-layer-shows-almost-no-structure-between-count-and-width-probesA random head in the same layer shows almost no structure between count and width probes
Control for twisting.
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Evidence for twisting.
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.
- Ringing evidence.
- Geometric evidence for convergence to stable truth directions only for simpler tasks.
- Will a head perform 'the same algorithm' across prompts, or are individual heads polysemantic?question0.773Head polysemanticity.
- Central theme.
- Argues against the single-layer analysis approach of prior work.
- Thought detection peaks at ~2/3 layer depth; intention checking peaks at ~1/2 layer depth.finding0.769Lindsey (2026) differential layer performance explained by Janus's path combinatorics — different tasks use different path distributions.
- Interpretation of eval loss result.
- Supported by the geometric transition visible in cosine similarity heatmaps for F0-F3.