finding
active
finding:shared-features-decoder-vectors-are-almost-always-highly-aligned-across-base-and-finetuned-model-with-a-few-thousand-showing-low-or-negative-correlationShared features' decoder vectors are almost always highly aligned across base and finetuned model, with a few thousand showing low or negative correlation
Same concept and function in both models mostly.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Claims (1)
claim
- Suspicion not investigated in depth.
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.
- Third empirical pattern.
- Larger rotations between shared features skew decoder cosine similarities downward in the toy modelfinding0.793Supports 'same features, used differently' interpretation.
- Footnote 2.
- Low-cosine shared features are best interpreted as same features used differently in the two modelsclaim0.782Interpretation of cosine distribution.
- Explicitly identified research gap: anecdotal evidence exists but rigorous characterization is absent
- Near-identical numbers of exclusive features are found for both models in base vs finetuned diffsfinding0.779Symmetry of exclusive features across models.
- Key limitation of the PRH for non-bijective observations
- Preliminary observation.