framework
active
framework:cross-layer-transcodersCross-Layer Transcoders
MLP replacement with sparse features underlying the original attribution graphs.
Neighborhood — ranked by edge-count
Papers (2)
paper
Methods (4)
method
- Gradient-based technique using SAE features to estimate causal effects on completions; used to corroborate NLA findings.
- TranscodersextendsDecomposition method for activations; VPD is compared against transcoders in sparsity-reconstruction tradeoff.
- CLT plus error nodes and original attention patterns, local to a prompt.
- Substitutes all MLP outputs with CLT reconstructions.
Frameworks (2)
framework
- Crosscodersrelated_toLearn a shared feature space across two models; basis of the work.
- Mask virtual weights during training targeting interference.
Artifacts (1)
artifact
- GPT-2 SmallaboutBase model on which CLTs are trained.
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.
- Used to express MLP in monosemantic features, 4096 features with JumpReLU.
- Replacement model architecture with 30M features replacing MLP neurons.
- Uses activations from one layer to predict the next; special case of crosscoders.
- Discussion.
- Decomposition approach mixing frozen base attention patterns.
- Cross-layer transcoder features form a mix of local, global and intermediate features by decoder normfinding0.807Preliminary decoder norm analysis.
- Each feature assigned a window of K layers to reduce training FLOPs.
- Translates steering vector via weighted average of shared decoder vectors.