The examples of features found in language models suggest they are highly sparse variables, consistent with dictionary learning being applicable

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Learned features reflect the functionality of the model and not just the data distribution, as evidenced by interpretable downstream effects not used in dictionary learningclaim0.842
  Authors argue features are model properties because logit effects and ablations are consistent with feature interpretations
• Sparse Autoencoders Find Highly Interpretable Features in Language Models (Cunningham et al., 2023)concept0.830
  Core methodology paper for SAE-based interpretable feature extraction
• as the subject model scales, how does the ideal expansion factor and required training data for dictionary learning change?question0.822
  Scaling laws for dictionary learning are unknown and needed to assess feasibility on frontier models
• Dictionary learning on model with randomly shuffled weights produces mainly single-token and poorly interpretable featuresfinding0.818
  Controls for dataset structure, showing trained model activations have richer structure than data distribution alone
• what metrics can reliably tell us if dictionary learning has successfully extracted high quality features?question0.812
  Central methodological gap: current metrics (loss, density histograms, manual inspection) are inadequate
• Sparse autoencoders are preferable to stronger iterative dictionary learning methods because they cannot recover features the model itself cannot accessclaim0.808
  Rationale for using simpler sparse autoencoders rather than NP-hard compressed sensing algorithms
• Large language models develop surprisingly coherent yet often rigid internal preferences as they scalefinding0.801
  Mazeika et al. finding reinforcing the need for emptiness-based flexible value architectures
• what is the 'correct number of features' for dictionary learning, and is this question well-posed?question0.801
  Open question about whether there is a true discrete feature count or a continuous splitting process