Jacobian Scopes: token-level causal attributions in LLMs

A team of researchers has released Jacobian Scopes, a set of gradient-based methods designed to trace how individual input tokens shape the predictions of large language models, according to a paper posted on the arXiv preprint server [1]. The work, authored by Jianbang Liu and colleagues, was first submitted to arXiv on 23 January 2026 and updated through 11 June 2026 [1]. The suite of tools is grounded in perturbation theory and information geometry, and it quantifies token-level influence on specific logits, the full predictive distribution, and a model's effective temperature [1][2]. Through case studies on instruction understanding, translation, and in-context learning, the authors show that Jacobian Scopes can surface implicit political biases and uncover word- and phrase-level translation strategies [1][2]. The paper also examines mechanisms behind in-context time-series forecasting, a topic the authors describe as recently debated [2]. The researchers have open-sourced their implementations and provide a cloud-hosted interactive demo on Hugging Face Spaces to let users explore Jacobian Scopes on custom text [2]. The demo is accessible through a link on the paper's abstract page, which is part of a broader ecosystem of community-built tools on arXiv [2][4]. arXiv, an open-access repository that began in 1991, now hosts more than two million e-prints and receives roughly 24,000 submissions per month [6]. The platform's arXivLabs framework, launched in 2020, allows third-party developers to integrate experimental features such as bibliographic explorers and code finders directly on article pages [4][5]. Large language models, which are trained on vast text corpora through self-supervised learning, make next-token predictions based on clues in their context [8]. The proliferation of layers and attention heads in modern architectures has made it difficult to determine which prior tokens most strongly influence a given prediction [1][2]. Jacobian Scopes addresses this gap by providing token-level causal attributions, a capability the authors argue can shed light on model behavior that is otherwise opaque [1][2].