claim

active

claim:a-good-parameter-subcomponent-is-causally-important-only-for-specific-roles-and-can-be-removed-from-the-model-without-hurting-performance-on-irrelevant-prompts

A good parameter subcomponent is causally important only for specific roles and can be removed from the model without hurting performance on irrelevant prompts

Definitional principle guiding VPD: subcomponents should encode narrow, targeted computational roles rather than distributed, multi-purpose functionality.

Source paper

extracted_from

cimcWhitepaper

Neighborhood — ranked by edge-count

Communities (4)

community

Mechanistic interpretability & model evaluation
members_of
Spans attention head decomposition, benchmark awareness, and genomic pathogenicity prediction via neural models.
Mechanistic interpretability via parameter decomposition
members_of
Tracing information flow through weight matrices and attention heads using attribution graphs to identify causally important subcomponents in language models.
Causal parameter decomposition in neural networks
members_of
Isolating interpretable, role-specific model subcomponents through causal analysis and targeted edits to understand mechanistic function.
Causal parameter subcomponent isolation
members_of
Identifying model components causally responsible for specific behaviors while removable for irrelevant tasks

Methods (1)

method

Adversarial Parameter Decomposition (VPD)
associated_with
Core technique introduced in this paper for decomposing neural network weight matrices into mechanistically simple, interpretable rank-one subcomponents.

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.

Parameter subcomponents cleanly isolate true mechanisms of the modelclaim0.811
Interpretive claim that the subcomponents correspond to real functional units.
Parameter subcomponentconcept0.784
One of the simple rank-one matrices resulting from VPD that sums with others to reconstruct the original model weights and has a specific functional role.
Investigating the causal substructure of neural representations is necessary to avoid misidentifying data structures of simpler representations as abstract conceptsclaim0.782
Motivated by the finding that lexical entailment decomposes into word identities.
Object models in which time, versioning, causality, etc., are significant are probably far better modelled by considering the time component as another key rather than an intrinsic property of the underlying model.claim0.772
Claim that orthogonal dimensions like time should be explicit keys in the associative model.
are the resulting parameter subcomponents actually interpretable objects?question0.761
First question posed after applying VPD, investigating whether the subcomponents make sense.
can we use these parameter subcomponents to perform clean, targeted changes?question0.760
Implicit question driving the editing experiment.
VPD subcomponents are sparse, interpretable, and avoid feature splitting.claim0.750
Assertion about the qualitative advantages of VPD's rank-one decomposition.
The strengths of all three models (concurrent OOP, logic, functional) are irrelevant to parallelism, and generally unhelpful in dealing with process creation and coordination.claim0.746
Assertion that the popular models add nothing to parallel programming.