claim
active
claim:key-query-and-value-vectors-are-intermediary-byproducts-w-ov-and-w-qk-are-the-fundamental-low-rank-matrices-describing-attention-head-behaviorKey, query, and value vectors are intermediary byproducts; W_OV and W_QK are the fundamental low-rank matrices describing attention head behavior
Reframing observation: the canonical K/Q/V decomposition is computationally convenient but not the most interpretable representation
Neighborhood — ranked by edge-count
Claims (1)
claim
- Key decomposition enabling separate analysis of where attention goes and what it does
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.
- What matrix decomposition or dimensionality reduction best summarizes the enormous low-rank OV and QK matrices?question0.771Open methodological question about converting the 50k x 50k expanded matrices into human-graspable summaries
- A pair of query and key subcomponents distributed across attention heads performs previous-token behaviorfinding0.770VPD recovers an attention algorithm for attending to the previous token, distributed across multiple heads.
- Evidence that core representations like preferences are persona-relative, supporting claim that personas gate content of representations
- Preliminary analyses suggest QK attribution matrices per head and position pair are approximately low-rankfinding0.767Motivates distillation.
- Basis-independent corroboration.
- Simple DoM vectors coherently represent reward hacking and are generalizable and interpretableclaim0.761Central claim.
- Fig. 13.
- Proposed explanation for why single-turn reformulation improves performance: models' training distribution is concentrated on single-turn reasoning.