hypothesis
active
hypothesis:a-better-basis-e-g-factoring-tokens-into-languages-or-parts-of-speech-may-yield-a-sparse-interpretable-modelA better basis (e.g., factoring tokens into languages or parts of speech) may yield a sparse interpretable model
Same model can be dense in one basis and sparse in another.
Source paper
extracted_fromNicholas L. Turner · Jeffrey Wu · Joshua Batson
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (2)
finding
- Parameter count roughly matches original transformer.
- Even restricted to most effective and helpful.
Claims (2)
claim
- Better bases may give sparsity.
- Discussion.
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.
- Sparse Autoencoders Find Highly Interpretable Features in Language Models (Cunningham et al., 2023)concept0.848Core methodology paper for SAE-based interpretable feature extraction
- Cited as enabling precise behavioral control through SAE features, extending the same methodological line
- Motivation for using sparsity-based dictionary learning on language models
- Sparse interpreter models may represent and decompose computational mechanisms differentlyclaim0.835Hinted by differences.
- Central claim of the paper: the method scales to state-of-the-art transformers.
- Practical implication.
- Articulates why a one-layer transformer with MLP is the appropriate starting target for mechanistic interpretability
- Forward-looking prediction about scalability of the method to larger models