method
active
method:head-loadingsHead Loadings
Per-edge contribution of each attention head to attribution graph edges.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Tracing Attention Computation Through Feature Interactionsintroducesmentions
Methods (2)
method
- Attribution GraphsextendsGradient-based technique using SAE features to estimate causal effects on completions; used to corroborate NLA findings.
- Strategy forcing edges to be mediated only by head paths of length 1 to avoid exponential path growth.
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.
- Modules loaded on demand at command invocation or through programmed calls; no separate linker; each module present once in memory.
- Quadratic scaling.
- Cosine similarity of a concept's lens vector with the residual stream; predicts swap success.
- The pass rate among a model's skill-loaded trajectories, measuring outcome conditioned on harness activation
- Capability to load application classes at runtime via late binding, enabling piecemeal addition of features without recompilation.
- Attention heads whose QK twists count manifolds to detect approaching line end.
- Transformer attention heads that could be recruited to extract different kinds of information (text vs. thoughts).
- The fraction of trajectories in which an agent actively loads at least one skill into its context