VPD (adVersarial Parameter Decomposition)

Core methodological framework introduced in this paper; decomposes weight matrices into rank-one interpretable subcomponents using adversarial ablations.

Neighborhood — ranked by edge-count

Papers (1)

paper

Interpreting Language Model Parameters
implements

Thinkers (8)

thinker

Lee Sharkey
introduces
Co-author of VPD paper; mechanistic interpretability researcher affiliated with Goodfire.
Bart Bussmann
introduces
Co-author of VPD paper; contributes to Goodfire mechanistic interpretability efforts.
Dan Braun
introduces
Co-author of VPD paper; contributes to mechanistic interpretability research at Goodfire.
Linda Linsefors
introduces
Co-author of VPD paper; contributes to parameter decomposition methodology.
Lucius Bushnaq
introduces
Lead author of VPD paper; researcher at Goodfire focused on mechanistic interpretability of language models.
Michael Ivanitskiy
introduces
Co-author of VPD paper; involved in mechanistic interpretability at Goodfire.
Nathan Hu
introduces
Co-author of VPD paper; participates in parameter decomposition research.
Oliver Clive-Griffin
introduces
Co-author of VPD paper; involved in parameter decomposition research.

Frameworks (1)

framework

Sparse Autoencoders (SAE) activation-based paradigm
contradicts
Standard interpretability approach that VPD critiques and proposes an alternative to.

Communities (1)

community

Neural Geometry
members_of

Methods (1)

method

Adversarial ablation
implements
Technique used in VPD to enforce mechanistic faithfulness of parameter decompositions.

Concepts (3)

concept

Parameter Decomposition (vs Activation)
aboutassociated_with
Manual model editing through parameter manipulation
about
Application enabled by VPD: direct manipulation of weight matrices for interpretable model modification.
SAE Critique (Manifold Shattering)
associated_with

Artifacts (1)

artifact

param-decomp (GitHub repository)
implements
Open-source implementation of VPD framework; code repository accompanying the research paper.

Related by similarity (8)

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.

Adversarial Parameter Decomposition (VPD)method0.957
Core technique introduced in this paper for decomposing neural network weight matrices into mechanistically simple, interpretable rank-one subcomponents.
VPD enables manual model editing through direct parameter manipulation.claim0.778
Applied capability claim: VPD enables surgical changes to model behaviour at the parameter level.
VPD decomposes parameters, not activations, flipping the standard SAE / activation-patching paradigm.claim0.774
Core proposition of the paper: a substrate-level critique of existing interpretability methods.
VPD subcomponents are sparse, interpretable, and avoid feature splitting.claim0.755
Assertion about the qualitative advantages of VPD's rank-one decomposition.
VPD identifies real, computational structure in neural network parametersclaim0.755
Central claim that VPD successfully uncovers genuine mechanisms.
Partial Information Decomposition (PID)framework0.753
Williams and Beer's decomposition of joint mutual information into unique, redundant, and synergistic components.
VPD scales to a 4-layer 67M-parameter model trained on The Pile.finding0.732
Empirical demonstration of VPD on a mid-scale transformer, establishing feasibility.
VPD can be arbitrarily applied to any neural network architectureclaim0.730
Claim of generality, highlighted as a key strength.