method
active
method:mlp-layer-null-interventionMLP-layer null intervention
Zeroing MLP layers to localize task neurons.
Neighborhood — ranked by edge-count
Papers (1)
paper
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.
- Feed-forward neural network with hidden layers, capable of representing non-linearly separable functions.
- Interpretation of pre/post MLP correlation.
- MLP neuron polysemanticity requires superposition, unlike heads.
- The fundamental operation of making in-place changes to model activations, placing the model in a counterfactual state
- Confirms attention rather than MLP as the locus of persona generation
- Key limitation of the paper's approach; MLP layers make up 2/3 of standard transformer parameters
- Fig. 10.
- Causal evidence that massive activations are required for stages of inference to emerge in looped models