Reproduce “Jacobian Scopes: A Unified Geometric Framework for Token-Level LLM Attribution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Toni J.B. Liu, Baran Zadeoğlu, Gurbir Arora, Raphaël Sarfati, Nicolas Boullé, Christopher J. Earls
Jacobian Scopes trace how input-token embeddings locally influence a language model’s next-token distribution. They use a vector–Jacobian product to score positions in one backward pass, projecting changes onto directions for a target logit, distributional sensitivity, or effective inverse temperature. Across two benchmarks, the best Scope beats non-Scope baselines in 10 of 12 settings; one is inconclusive, and Gemma-3 4B loses.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.