concept
active
concept:cross-architecture-model-diffingCross-architecture model diffing
Extension of model diffing to models with different architectures.
Neighborhood — ranked by edge-count
Papers (1)
paper
Concepts (1)
concept
- Model DiffingextendsApproach to identify interpretable differences between a base LLM and its fine-tuned version using SAE latents
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.
- Train a single SAE encoding and decoding activations from two models simultaneously to find shared and exclusive features.
- Whether learned cones transfer effectively across model families (Qwen vs Gemma) and sizes
- Conclusion.
- Synthetic two-model activations built from shared and exclusive latent factors with controllable ground truth.
- The paper's primary mechanistic analysis method: comparing SAE latent activations before and after fine-tuning to identify misalignment-relevant features
- Crosscoder model diffing is best used as a high-recall pre-screening tool followed by verification.claim0.754Recommended workflow.
- Injecting the activation difference between fine-tuned and original model into the AO.
- Ground-truth concept environment with random affine transform for Model B.