method
active
method:sparse-crosscoderSparse Crosscoder
Sparse dictionary learning variant that reads and writes to multiple layers or models, yielding shared features.
Neighborhood — ranked by edge-count
Frameworks (2)
framework
- Superposition HypothesisstudiesCore theoretical framework: neural networks represent more features than neurons by encoding features as directions in superposition
- Sparse AutoencoderextendsInterpretability framework used to decompose layer-40 activations into sparse feature sets for studying emotional alignment and persistence
Concepts (3)
concept
- Model DiffingimplementsApproach to identify interpretable differences between a base LLM and its fine-tuned version using SAE latents
- Cross-layer superpositionstudiesRepresentation of features spread across multiple layers, complicating dictionary learning.
- Removing duplicate persistent features to shrink circuit depth.
Methods (7)
method
- Source of the features used in attributions.
- Cross-layer transcoder (CLT)extendsReplacement model architecture with 30M features replacing MLP neurons.
- TranscoderextendsUses activations from one layer to predict the next; special case of crosscoders.
- Sparsity penalty weighting by sum of per-layer decoder norms, enabling comparability with per-layer SAEs.
- Acausal crosscoderextendsBasic crosscoder reading and reconstructing all layers jointly.
- Each feature assigned a window of K layers to reduce training FLOPs.
- Pre/post MLP crosscoderextendsEncoders read pre-MLP space; decoders reconstruct pre-MLP and MLP output to find newly computed features.
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.
- Extension of SAEs that jointly learns latents across representations from different models; proposed as alternative for extended fine-tuning
- Learn a shared feature space across two models; basis of the work.
- Coding scheme where qualities are represented by few neurons with continuous similarity relations.
- Used in Anthropic welfare assessment to identify performative behavior and hidden emotional struggle co-activating features
- Mask virtual weights during training targeting interference.
- Primary method introduced: trains a one-hidden-layer MLP with L1 sparsity penalty to decompose model activations into overcomplete feature dictionaries
- Candidate approach to solving superposition.
- The central mechanistic interpretability tool applied across all three EEG transformers to extract sparse feature dictionaries