finding
active
finding:pca-of-answer-selection-head-q-k-o-forms-one-continuous-cloud-with-no-clustersPCA of answer selection head Q, K, O forms one continuous cloud with no clusters
PCA uninformative here.
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- Best guess but not fully convinced; lacks crisp text description to test.
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.
- Confirms three behaviors.
- Used to analyze emotion vector geometry.
- Hypothesis-generation tool for detecting distinct head behaviors via clustering.
- Evidence that manifolds exist in LLM representations.
- Shows absence of abstract truth representations in smallest model, supporting scale-dependent emergence claim
- The Primer architecture's depthwise convolution change would allow induction heads to form without requiring K-compositionhypothesis0.710Architectural interpretation of how Primer's design change relates to the paper's mechanistic theory of induction heads
- Primary visual evidence for linear truth representations in large LLMs
- Limitations of PCA.