· Updated
DFC), a mechanistic interpretability tool that applies the software "diff" principle to neural networks. Unlike standard crosscoders, the DFC uses a shared dictionary for common concepts and dedicated sections for features exclusive to each model, identifying unique traits across different architectures.Traditional safety evaluations are reactive, testing only for known risks. This approach identifies "unknown unknowns"—emergent behaviors that differ between models—by analyzing internal "switches" rather than just outputs. It enables efficient auditing by focusing human review on the specific features that drive behavioral divergence between two different models.
Use this framework to understand the inherent biases of open-weight models like Llama-3.1-8B-Instruct or Qwen3-8B. Researchers found specific features controlling behaviors like American exceptionalism and political alignment that can be toggled. The full research paper and methodology are available for those building or auditing specialized model deployments.
New Anthropic Fellows Research: a new method for surfacing behavioral differences between AI models. We apply the “diff” principle from software development to compare open-weight AI models and identify features unique to each. Read more: https://t.co/VAsu2PSgCX
