finding
active
finding:decomposition-of-all-24-weight-matrices-in-a-67m-parameter-lm-yields-10-000-parameter-subcomponentsDecomposition of all 24 weight matrices in a 67M-parameter LM yields ~10,000 parameter subcomponents
Quantitative result of VPD application; the network's 24 matrices decompose into approximately 10,000 rank-one subcomponents.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Claims (1)
claim
- Central claim that VPD successfully uncovers genuine mechanisms.
Communities (3)
community
- Spans attention head decomposition, benchmark awareness, and genomic pathogenicity prediction via neural models.
- Tracing information flow through weight matrices and attention heads using attribution graphs to identify causally important subcomponents in language models.
- Tracing information flow through parameter subcomponents to isolate computational mechanisms for specific model predictions, using tools like attribution graphs and 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.
- SPD splits W_out into 50 subcomponents forming a single rank-50 component in compressed computation modelfinding0.740Contradicts APD's 100 components for W_out.
- Ground truth mechanisms identified.
- The core idea of decomposing weight matrices into components for interpretability.
- Figure 1.
- SPD decomposes W_in of compressed computation model into 100 subcomponents, one per input featurefinding0.728Neuron contributions match target.
- Concurrent work result showing emergent misalignment occurs in small models
- The layer approximately two-thirds through an LLM's transformer stack, reported to best predict human brain activity; identified as promising for consciousness indicators.
- Fig. 22.