Jacobian Scopes: token-level causal attributions in LLMs

Researchers introduce Jacobian Scopes, a new gradient-based method for interpreting how individual tokens influence Large Language Model predictions. The technique uses perturbation theory and information geometry to reveal model biases, translation strategies, and learning mechanisms, with open-source implementations and an interactive demo available.

• →Jacobian Scopes provides token-level causal attribution to understand which input tokens most influence LLM predictions.
• →The method can uncover implicit political biases and reveal word-level translation strategies in language models.
• →The technique sheds light on mechanisms underlying in-context learning and time-series forecasting capabilities.
• →Researchers have open-sourced their implementations and provided a cloud-hosted interactive demo for testing.
• →The approach is grounded in perturbation theory and information geometry to quantify token influence on predictions.