claim
active
claim:the-weak-linear-representation-hypothesis-is-supported-by-probes-steering-and-saes-while-the-strong-version-is-false-for-some-modelsThe weak linear representation hypothesis is supported by probes, steering and SAEs, while the strong version is false for some models
Position on LRH.
Source paper
extracted_from(2025) · Lee Sharkey · Bilal I. Chughtai · Joshua Batson · Jack W. Lindsey +25
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- A small transformer solved modular addition via a Fourier transform strategy understood only in retrospectassociated_withAlien algorithms example.
Frameworks (1)
framework
- The hypothesis that models internalize concepts as approximately linear directions in representation space; used to interpret MDS injection behavior
Claims (1)
claim
- Steering derived from LRH.
Questions (1)
question
- Open question A.1.1a.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.
- Authors' overall conclusion from number of interpretable features, activation-level correspondence to intensity, sensible logit weights, and interference weights
- Activations are sums of sparse features living in subspaces.
- Interpretive synthesis of DIM and cone intervention successes
- Hypothesis that information may be encoded in arbitrary non-linear subspaces of a neural network
- Author's interpretation of the VTAB alignment results echoing Tolstoy
- Establishes that the observed linear structure is not merely a representation of text probability
- The paper's critique of the standard linear steering baseline, supported by the days-of-week demo.
- Theoretical open question about the geometry of truth in LLMs raised in Discussion