hypothesis
active
hypothesis:positional-signal-is-diffuse-because-the-model-encodes-group-positions-in-a-poorly-linearly-separable-way-for-middle-groupsPositional signal is diffuse because the model encodes group positions in a poorly linearly separable way for middle groups
Tested via probes.
Source paper
extracted_from(2025) · Yoav Gur-Arieh · Mor Geva · Atticus Geiger
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Hidden states of middle entity groups become less linearly separable by position as n growsassociated_withsupportsFig. 22.
Methods (1)
method
- PCA and multinomial logistic regression probesassociated_withUsed to test linear separability of group position.
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 and reflexive mechanisms provide sharper signals refining the diffuse positional signalclaim0.824Takeaway.
- Challenges prevailing view.
- Central claim motivating DAS over prior methods.
- Finding (2).
- Plausible explanation.
- Neural Representations of Location Composed of Spatially Periodic Bands (Krupic et al., 2012)concept0.724Discovery of band cells; TEM-t also recapitulates these representations.
- Key quote connecting path redundancy to interferometric information encoding.
- Key limitation of the PRH for non-bijective observations