finding
active
finding:20m-model-non-task-gradient-has-cosine-similarity-0-10-0-09-with-task-direction-with-wildly-oscillating-similarity20M model non-task gradient has cosine similarity 0.10±0.09 with task direction, with wildly oscillating similarity
Fig. 9.
Source paper
extracted_from(2026) · Jing Huang · Daniel Wurgaft · Rachit Bansal · Laura Ruis +5
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- Hypothesis box in §3.2.
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.
- Fig. 9.
- 1B model batch-gradient similarity to task reference is 0.08±0.02 at injection steps; 300M is 0.04±0.04finding0.808Fig. 9.
- Experiment 4 result showing DIM captures only one facet of the multi-dimensional truth subspace
- Experiment 3 comparison: zero-shot control shows lower semantic convergence than experimental condition
- Appendix E replication of DIM alignment finding in Qwen model
- High cosine similarity for Gemma3 steering vectors suggests strong linear reflection structure.
- Core result of Experiment 3: cross-model semantic convergence under self-referential processing
- Probe cosine similarity becomes more peaked with deeper layers 0-3, with higher secondary ringsfinding0.767Sharpening.