concept
active
concept:datapoint-featureDatapoint feature
Transcoder feature that activates on p and produces p, memorizing something the model did not.
Neighborhood — ranked by edge-count
Papers (1)
paper
- A Toy Model of Mechanistic (Un)Faithfulnesscitesmentions
Methods (2)
method
- Crude measure of excess explanation of p over constituent dimensions, summed across features.
- Training distribution where special point p=[1,1,1,0,...] appears with fraction repeat_frac.
Concepts (1)
concept
- Mechanistic faithfulnessassociated_withIdea that decomposed components must causally reflect the model's computation; enforced by adversarial ablation in VPD.
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.
- Mitigation technique that filters out datapoints identified by probe-based ranking.
- Individual examples used during post-training that can cause specific behaviors.
- Datapoint features form only if there is a repeated datapoint in training and sparsity regularization is appliedhypothesis0.764Two necessary conditions.
- Interpretable concept representations hypothesized as basic units of computation.
- Property that features activate on only a small fraction of inputs; enables compressed sensing and is what allows superposition to work
- Method of optimizing input to cause a neuron to fire maximally, used to characterize what a neuron detects; establishes causal link
- Features that activate before the action sequence they bring about, fulfilling functions of intention
- Conventional programming constructs like variables, arrays; claimed unnecessary for Elephant programs.