finding
active
finding:country-query-vs-own-city-key-cosine-similarity-mean-0-47-versus-0-01-for-other-citiesCountry query vs own-city key cosine similarity mean −0.47, versus +0.01 for other cities
QK geometry anti-alignment.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Supported by geometry and 98% result but not a full account.
Methods (1)
method
- Measures QK geometry between country queries and city keys.
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.
- In the residual stream a country and its own cities have positive cosine similarity (mean +0.17)finding0.795Head flips sign via W_Q and W_K.
- Appendix E replication of DIM alignment finding in Qwen model
- Distance matrices of My and Mz correlate r=.71 (p<.001) for Arbitrary domain, lower similarityfinding0.748Weaker shared structure.
- Core result of Experiment 3: cross-model semantic convergence under self-referential processing
- Experiment 3 comparison: zero-shot control shows lower semantic convergence than experimental condition
- Used to measure alignment between DIM direction and cone basis vectors to assess overlap
- High cosine similarity for Gemma3 steering vectors suggests strong linear reflection structure.
- Gemma Scope decoder nearest-neighbor cosine similarities skewed above random but few extremely highfinding0.734Fig 7.