Glossary · term

Cross-Architecture Model Diffing with Crosscoders

The use of crosscoders to compare models with different architectures and detect behavioral differences without a base-to-finetune relationship. This unsupervised method extracts features that distinguish models, for example alignment with the CCP party line in Qwen3-8B or "american exceptionalism" in Llama-3.1-8B. It matters because new releases are usually new architectures rather than simple fine-tunes.

Safety2025Wave 3 · 2025–26Maturity: 1/5

Maturity rationale

speculative / early neologism (warning)

References

Author: Trenton Bricken