This paper introduces a method for automatically interpreting language models by analyzing information flow routes. The approach builds graphs representing token representations and computations within the model, extracting important subgraphs through top-down tracing and attribution, rather than activation patching. This method is more efficient, versatile, and informative than existing techniques, allowing analysis of model behavior for various predictions and domains. Experiments on Llama 2 show that certain attention heads, such as previous token and subword merging heads, are crucial for tasks like Indirect Object Identification and Greater-than. The method also reveals domain-specific model components, such as those specialized for coding or multilingual texts. Compared to patching, the information flow approach is about 100 times faster and provides more comprehensive insights into model behavior. The study further demonstrates that information flow routes can reveal patterns based on token parts of speech, subword positions, and domain-specific functions. Overall, the method offers a powerful tool for understanding and interpreting large language models.This paper introduces a method for automatically interpreting language models by analyzing information flow routes. The approach builds graphs representing token representations and computations within the model, extracting important subgraphs through top-down tracing and attribution, rather than activation patching. This method is more efficient, versatile, and informative than existing techniques, allowing analysis of model behavior for various predictions and domains. Experiments on Llama 2 show that certain attention heads, such as previous token and subword merging heads, are crucial for tasks like Indirect Object Identification and Greater-than. The method also reveals domain-specific model components, such as those specialized for coding or multilingual texts. Compared to patching, the information flow approach is about 100 times faster and provides more comprehensive insights into model behavior. The study further demonstrates that information flow routes can reveal patterns based on token parts of speech, subword positions, and domain-specific functions. Overall, the method offers a powerful tool for understanding and interpreting large language models.