method
active
method:model-diffing-with-sparse-autoencodersModel-Diffing with Sparse Autoencoders
The paper's primary mechanistic analysis method: comparing SAE latent activations before and after fine-tuning to identify misalignment-relevant features
Neighborhood — ranked by edge-count
Papers (1)
paper
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.
- Approach to identify interpretable differences between a base LLM and its fine-tuned version using SAE latents
- Train a single SAE encoding and decoding activations from two models simultaneously to find shared and exclusive features.
- Interpretability framework used to decompose layer-40 activations into sparse feature sets for studying emotional alignment and persistence
- Interpretability method criticized in this paper for shattering manifolds into atomic pieces, obscuring overarching semantic structure.
- Conclusion.
- Used in Anthropic welfare assessment to identify performative behavior and hidden emotional struggle co-activating features
- Interpretability tool underlying top-feature, details, steering tools.
- Central claim of the paper: the method scales to state-of-the-art transformers.