finding
active
finding:in-the-boundary-head-s-qk-space-character-count-and-line-width-probes-reach-max-cosine-similarity-1-at-an-offset-versus-0-25-in-the-residual-streamIn the boundary head's QK space, character count and line width probes reach max cosine similarity ≈1 at an offset, versus ~0.25 in the residual stream
Evidence for twisting.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Role of extra dimensions.
Questions (1)
question
- How does the model detect the boundary?answered_byIntro question.
Findings (1)
finding
- Control for twisting.
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.
- On DINOv3, classification and segmentation tolerate R²≈0.8, while depth needs R²≈0.9 to retain 95% accuracy.finding0.797Task-dependent distortion floors.
- Mechanistic evidence that network actively attenuates injected perturbations, explaining late-layer introspection failure
- Probe cosine similarity becomes more peaked with deeper layers 0-3, with higher secondary ringsfinding0.790Sharpening.
- In the residual stream a country and its own cities have positive cosine similarity (mean +0.17)finding0.789Head flips sign via W_Q and W_K.
- Appendix E replication of DIM alignment finding in Qwen model
- Core result of Experiment 3: cross-model semantic convergence under self-referential processing
- First two PCs of characters remaining space capture 92% of variance; summed head outputs form a 2D curvefinding0.782Sum of near-1D head outputs.
- Linear readout works but is not the true representation.