claim
active
claim:fully-decoupled-saes-are-causally-disconnected-and-cannot-support-attributionFully decoupled SAEs are causally disconnected and cannot support attribution
Per-token SAE input is invariant to uniform shifts from ablating turn-averaged features.
Source paper
extracted_from(2026) · Kevin Der · Harish Kamath · Thompson, Ben
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Table 1.
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.
- Hybrid with separate turn-averaged and residual per-token SAEs.
- Computing attribution as the dot product of the output logit gradient with the SAE decoder weight, multiplied by feature activation.
- SAEs may have the right inductive biases for deeper layers but not middle or earlier layers.claim0.752Speculation about layer dependence.
- Limitation statement.
- Claim that feature grounding enables interpretability metrics.
- Uncertain α and β.
- Geometric explanation for prior negative results.
- Core proposition of the paper: a substrate-level critique of existing interpretability methods.