method
active
method:weakly-causal-crosscoder-10m-featuresWeakly causal crosscoder (10M features)
Source of the features used in attributions.
Neighborhood — ranked by edge-count
Papers (2)
paper
- HeadVis: An Interactive Tool For Investigating Attention Headsintroducesmentions
Methods (4)
method
- Dictionary used for all main-text features.
- Sparse CrosscoderextendsSparse dictionary learning variant that reads and writes to multiple layers or models, yielding shared features.
- Writes attention scores/head output as sums of feature interactions; core HeadVis capability.
- Third checkpointing strategy; adopted here.
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.
- Naive implementation results.
- Decoder cost dominates.
- Extension of SAEs that jointly learns latents across representations from different models; proposed as alternative for extended fine-tuning
- Motivation for DFC.
- Decoder norm profile across 18 layers.
- Train a single SAE encoding and decoding activations from two models simultaneously to find shared and exclusive features.
- Family of count features.
- Circuit simplification theory.