Superscopes: Amplifying Internal Feature Representations for Language Model Interpretation
This work addresses the problem of mechanistic interpretability for researchers and practitioners working with LLMs, representing an incremental improvement over existing methods like Patchscopes.
The paper tackles the challenge of interpreting internal representations in large language models by introducing Superscopes, a technique that amplifies superposed features in MLP outputs and hidden states before patching them into new contexts, enabling interpretation of previously unexplained representations without additional training.
Understanding and interpreting the internal representations of large language models (LLMs) remains an open challenge. Patchscopes introduced a method for probing internal activations by patching them into new prompts, prompting models to self-explain their hidden representations. We introduce Superscopes, a technique that systematically amplifies superposed features in MLP outputs (multilayer perceptron) and hidden states before patching them into new contexts. Inspired by the "features as directions" perspective and the Classifier-Free Guidance (CFG) approach from diffusion models, Superscopes amplifies weak but meaningful features, enabling the interpretation of internal representations that previous methods failed to explain-all without requiring additional training. This approach provides new insights into how LLMs build context and represent complex concepts, further advancing mechanistic interpretability.