finding
active
finding:layer-9-is-approximately-optimal-for-linear-probes-across-all-domainsLayer 9 is approximately optimal for linear probes across all domains
Layer selection.
Source paper
extracted_from(2026) · Eric Bigelow · Raphaël Sarfati · Daniel Wurgaft · Owen Lewis +4
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Beliefs linearly decodable.
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.
- Argues against the single-layer analysis approach of prior work.
- Self-repair effect.
- E3 result establishing the Goldilocks zone at mid-layers for LLaMA architecture
- Probing limited.
- Supported by the geometric transition visible in cosine similarity heatmaps for F0-F3.
- Intrinsic noise in representation.
- Linear probe achieves 100% classification accuracy for almost all components in Pythia-6.9B gender taskfinding0.753Demonstrates that linear probes can overestimate causal relevance; probes succeed on non-causally-relevant representations
- Supports the claim against single-layer probing approaches used in prior work.