Ridge Regression on Message Embeddings

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

• Ridge Regression Probingmethod0.793
  Ridge regression fit on top-256 PCs of Gemini embeddings to predict model layer-40 activations and compute residuals
• Ridge regression probe constructionmethod0.784
  Method used to predict model activations from Gemini embeddings and compute residuals for probe construction
• User message embeddings predict subsequent model Assistant Axis projection with R2=0.53-0.77 (p<0.001) but predict delta from previous response with only R2=0.10finding0.734
  Shows model persona position is primarily determined by the most recent user message, not prior drift
• Input Embedding Similarity Baselinemethod0.705
  Baseline method for instruction discovery using surface-level input embedding similarity instead of steering vectors.
• Message Passing Inferenceconcept0.701
  Algorithmic framework for probabilistic inference in graphical models.
• span embedding analysismethod0.695
  Extracting embeddings from instruction and example spans.
• Sparse Autoencoders Find Highly Interpretable Features in Language Models (Cunningham et al., 2023)concept0.693
  Core methodology paper for SAE-based interpretable feature extraction
• learnable hidden embeddingsconcept0.690
  The component used in latent reasoning to perform internal computation.