finding
active
finding:learned-positional-widens-for-middle-indices-and-narrows-toward-the-endLearned positional σ widens for middle indices and narrows toward the end
Learned parameters.
Source paper
extracted_from(2025) · Yoav Gur-Arieh · Mor Geva · Atticus Geiger
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Challenges prevailing view.
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.
- Lexical contribution amplified near positional index but suppressed near reflexive indexfinding0.778Interaction patterns in Figure 3.
- Mixed predictions cluster near positional index with diffuse logit distribution in middle indicesfinding0.775Figures 3, 13.
- Figures 14–17.
- Forward-looking interpretive claim about the implications of recurrent position encodings for NLP research.
- Interpretation of padding results.
- From Eq. 5.
- Novel interpretive claim about position encodings inspired by the TEM-transformer correspondence.
- Learned rotations reveal that direct probes over standard activation bases would miss the actual causal role of representations.