Paper 2602.11729
Cross-Architecture Model Diffing with Crosscoders: Unsupervised Discovery of Differences Between LLMs
- Published
- Feb 2026
- Research lab
- Anthropic
- Citations
- 9
- GitHub
- Not linked
01 In brief
Summary
This paper presents the first application of crosscoders to cross-architecture model diffing, introducing Dedicated Feature Crosscoders (DFCs) to better isolate model-exclusive features.
DFCs partition the feature space into model-exclusive and shared sets, overcoming the standard crosscoder's prior toward shared features.
In synthetic toy models, DFCs recover more true exclusive concepts at the cost of more false positives, a favorable trade-off for safety auditing.
On real models, DFCs identify features with higher exclusivity scores and uncover meaningful behavioral differences: CCP alignment in Qwen3-8B and Deepseek-R1-0528-Qwen3-8B, American exceptionalism in Llama3.1-8B-Instruct, and a copyright refusal mechanism in GPT-OSS-20B.
Steering these features causally shifts model outputs, and cross-architecture transfer of persona vectors validates the learned alignment.
The method is proposed as a high-recall pre-screening tool for discovering unknown unknowns, with limitations including sensitivity to training initialization and the need for further validation.
02 From the paper
Abstract
Model diffing, the process of comparing models' internal representations to identify their differences, is a promising approach for uncovering safety-critical behaviors in new models. However, its application has so far been primarily focused on comparing a base model with its finetune. Since new LLM releases are often novel architectures, cross-architecture methods are essential to make model diffing widely applicable. Crosscoders are one solution capable of cross-architecture model diffing but have only ever been applied to base vs finetune comparisons. We provide the first application of crosscoders to cross-architecture model diffing and introduce Dedicated Feature Crosscoders (DFCs), an architectural modification designed to better isolate features unique to one model. Using this technique, we find in an unsupervised fashion features including Chinese Communist Party alignment in Qwen3-8B and Deepseek-R1-0528-Qwen3-8B, American exceptionalism in Llama3.1-8B-Instruct, and a copyright refusal mechanism in GPT-OSS-20B. Together, our results work towards establishing cross-architecture crosscoder model diffing as an effective method for identifying meaningful behavioral differences between AI models.