Fisher Scopes

Fisher Scopes is a specific variant within the Jacobian Scopes framework, a suite of gradient-based, token-level causal attribution methods developed for interpreting predictions made by large language models (LLMs). Its precise technical definition is a method that quantifies the sensitivity of the *full predictive distribution* with respect to input tokens. This is achieved by analyzing the linearized relations between the final hidden state and the inputs, effectively measuring how each prior token causally influences the entire probability distribution over the next token. Fisher Scopes matters because it addresses the challenge of understanding which parts of the input context most strongly influence an LLM's complex predictions, especially given the deep and intricate architectures of modern models. By providing fine-grained insights into the causal impact of input tokens, it helps researchers and ML engineers diagnose model behavior, uncover biases, and shed light on mechanisms like in-context learning. It is primarily used in LLM interpretability research, bias detection, and understanding complex model behaviors in applications like instruction understanding, translation, and time-series forecasting.