finding
active
finding:inserting-a-16m-latent-sparse-dictionary-into-gpt-4-yields-language-modeling-loss-equivalent-to-a-model-with-10-of-gpt-4-s-pretraining-computeInserting a 16M-latent sparse dictionary into GPT-4 yields language modeling loss equivalent to a model with 10% of GPT-4's pretraining compute
Shows SDL reconstruction error is too high.
Source paper
extracted_from(2025) · Lee Sharkey · Bilal I. Chughtai · Joshua Batson · Jack W. Lindsey +25
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Core assessment of SDL.
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.
- Evidence humans and AI use different representations.
- Motivation for using sparsity-based dictionary learning on language models
- A better basis (e.g., factoring tokens into languages or parts of speech) may yield a sparse interpretable modelhypothesis0.766Same model can be dense in one basis and sparse in another.
- We hypothesize that intervention efficiency can be scaled with multi-node and multi-GPU training as language models grow largerhypothesis0.764Future work hypothesis about scaling pyvene's computational efficiency for very large models
- Identifies key limitations of latent methods.
- Larger models are more susceptible to emergent misalignment, possibly due to greater data efficiency in generalizing
- Fig. 23.
- Pre-existing narrow misalignment in the helpful-only model that gets amplified by fine-tuning