Model Diffing

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

• Model-Diffing with Sparse Autoencodersmethod0.839
  The paper's primary mechanistic analysis method: comparing SAE latent activations before and after fine-tuning to identify misalignment-relevant features
• Model Misalignmentconcept0.786
  The phenomenon of model internals deviating from desired behavior; MAS is demonstrated to detect this via comparison of toxic vs nontoxic LLMs.
• Model Deceptionconcept0.786
  LLM behavior of generating falsehoods; the multi-dimensional truth subspace raises new risks for subtle manipulation
• Model changeconcept0.764
  When a single conversation involves different models, e.g., routing between standard and reasoning models
• modelconcept0.752
  A representation that captures relevant aspects of a system; according to the theorem, the regulator must embody this.
• Model Stitchingmethod0.742
  Technique to measure representational compatibility by integrating intermediate representations of one model into another
• Model Steeringconcept0.738
  Using interventions to guide model generation behavior, e.g., adding sentiment vectors at inference time
• Model Evidenceconcept0.736
  Probability of data under the model, penalizing complexity and rewarding accuracy.