method
active
method:crosscoder-model-diffingCrosscoder model diffing
Train a single SAE encoding and decoding activations from two models simultaneously to find shared and exclusive features.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Insights on Crosscoder Model Diffingintroducesmentions
Concepts (1)
concept
- Model DiffingextendsApproach to identify interpretable differences between a base LLM and its fine-tuned version using SAE latents
Methods (3)
method
- Toy model of crosscoder diffingrelated_toSynthetic two-model activations built from shared and exclusive latent factors with controllable ground truth.
- Key design choice encouraging feature exclusivity.
- Proposed variation designating ~10k of ~250k features as shared with penalty 0.1-0.2x baseline to soak up shared variance.
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.
- Conclusion.
- Extension of model diffing to models with different architectures.
- Crosscoder model diffing is best used as a high-recall pre-screening tool followed by verification.claim0.859Recommended workflow.
- The paper's primary mechanistic analysis method: comparing SAE latent activations before and after fine-tuning to identify misalignment-relevant features
- A priori unknown behavior discovered.
- Learn a shared feature space across two models; basis of the work.
- Decoder norm profile across 18 layers.
- Circuit simplification theory.