claim
active
claim:datapoint-features-are-not-mechanistically-faithful-when-the-underlying-model-has-no-special-notion-of-pDatapoint features are not mechanistically faithful when the underlying model has no special notion of p
Memorization by transcoder not in model.
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Demonstrates unfaithfulness in a simple setup.
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.
- Datapoint features form only if there is a repeated datapoint in training and sparsity regularization is appliedhypothesis0.779Two necessary conditions.
- Limitation.
- Future work.
- Hyperparameter scan result.
- Motivates Jacobian connection.
- Authors' prioritization.
- While NLA claims can be false in specifics, they are typically thematically faithful to contextclaim0.749Key insight about confabulation patterns in NLAs enabling practical use.
- Takeaway of toy experiments.