framework
active
framework:crosscodersCrosscoders
Learn a shared feature space across two models; basis of the work.
Neighborhood — ranked by edge-count
Papers (3)
paper
Methods (2)
method
- Sparse Dictionary LearningextendsGeneral method for finding overcomplete sparse decompositions; the paper uses sparse autoencoders as an approximation
- Post-hoc metric for model exclusivity in standard crosscoders.
Frameworks (4)
framework
- Cross-Layer Transcodersrelated_toMLP replacement with sparse features underlying the original attribution graphs.
- Sparse Crosscodersrelated_toExtension of SAEs that jointly learns latents across representations from different models; proposed as alternative for extended fine-tuning
- Architectural variant partitioning the dictionary into A-exclusive, B-exclusive and shared sets.
- Baseline from Mishra-Sharma et al. dedicating features to shared variance.
Hypotheses (1)
hypothesis
- Proposed remedy.
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.
- Decomposition method for activations; VPD is compared against transcoders in sparsity-reconstruction tradeoff.
- Uses activations from one layer to predict the next; special case of crosscoders.
- Sparse dictionary learning variant that reads and writes to multiple layers or models, yielding shared features.
- Basic crosscoder reading and reconstructing all layers jointly.
- Mask virtual weights during training targeting interference.
- Used to express MLP in monosemantic features, 4096 features with JumpReLU.
- Each feature assigned a window of K layers to reduce training FLOPs.
- Train a single SAE encoding and decoding activations from two models simultaneously to find shared and exclusive features.