method
active
method:checkpointing-attention-paths-with-featuresCheckpointing Attention Paths with Features
Strategy forcing edges to be mediated only by head paths of length 1 to avoid exponential path growth.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Tracing Attention Computation Through Feature Interactionsintroducesmentions
Frameworks (1)
framework
- Decomposition approach mixing frozen base attention patterns.
Methods (3)
method
- Head LoadingsusesPer-edge contribution of each attention head to attribution graph edges.
- Third checkpointing strategy; adopted here.
- Second checkpointing strategy.
Claims (1)
claim
- Number of attention head paths between transcoder features grows exponentially with layer separationsupportsWhy checkpointing is needed.
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.
- Proposed attention analog of sparse monosemantic features for decomposing attention layers.
- Term introduced by this paper for the horizontal axis of information flow in transformers across token-positions via attention heads
- Process using Q, K, V to compute a heat map over K and weighted sum of V.
- Reason for adopting the third strategy.
- Features whose QK/OV are high rank, hard to interpret.
- Expectation for decomposition.
- Suspected but no confirmed example in real LLM.