Noisy Simulation of Sparse Networks

Mechanism by which superposition works: small neural networks exploit sparsity to approximately simulate much larger sparse networks

Neighborhood — ranked by edge-count

framework

Superposition Hypothesis
associated_with
Core theoretical framework: neural networks represent more features than neurons by encoding features as directions in superposition

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Sparse Coding Hypothesisconcept0.783
The hypothesis that semantic concepts in neural representations can be captured using sparsity priors
Sparse interpretability in neural networksconcept0.763
VPD achieves sparse, interpretable parameter subcomponents with improved sparsity-reconstruction tradeoff.
Sparse circuitsconcept0.755
A goal in mechanistic interpretability to identify sparse computational subgraphs; VPD promotes sparse parameter circuits.
Greedy Algorithm for Network Coarse-Grainingmethod0.743
Method to aggregate nodes in complex networks to maximize EI, proposed by Klein & Hoel.
Sparse and smooth codingconcept0.733
Coding scheme where qualities are represented by few neurons with continuous similarity relations.
Model-Diffing with Sparse Autoencodersmethod0.729
The paper's primary mechanistic analysis method: comparing SAE latent activations before and after fine-tuning to identify misalignment-relevant features
Sparse Dictionary Learningmethod0.728
General method for finding overcomplete sparse decompositions; the paper uses sparse autoencoders as an approximation
Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models (Marks et al., 2025)concept0.724
Cited as enabling precise behavioral control through SAE features, extending the same methodological line