Research map / Large language models / Language, safety and interpretability
Jacobian Scopes: A Unified Geometric Framework for Token-Level LLM Attributions
Toni Liu, Baran Zadeoğlu, Nicolas Boulle, Raphaël Sarfati, Gurbir Arora, Christopher Earls
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Jacobian Scopes, a gradient-based attribution framework that explains full predictive distributions of LLMs at the token level rather than only a single target logit.
Problem
Gradient-based attribution for LLMs has been limited to explaining scalar outputs (the logit of one target token), whereas questions about uncertainty and the spread of plausible continuations concern the full predictive distribution.
Approach
Extends gradient-based token attribution from scalar logits to the full predictive distribution via a unified geometric (Jacobian-based) framework, so that attributions can answer distribution-level questions such as which tokens drive model uncertainty or a broad set of plausible continuations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| attribute the full predictive distribution (Fisher Scope) instead of a single target logit (Input x Gradient) | LLaMA-3.2 1B · LAMBADA | AOPC: -1.28 -> -1.32 | Improved |
| attribute effective temperature (Temperature Scope) instead of Input x Gradient | Qwen2.5-1.5B · LAMBADA | AOPC: -1.72 -> -1.86 | Improved |
| attribute effective temperature (Temperature Scope) instead of Input x Gradient | Qwen2.5 3B · LAMBADA | AOPC: -1.39 -> -1.55 | Improved |
| use Fisher Scope instead of Integrated Gradients | Gemma-3 4B · IWSLT2017 DE-EN | AOPC: -1.34 -> -1.20 | Hurt |
| use Semantic Scope instead of Input x Gradient | LLaMA-3.2 1B · LAMBADA | AOPC: -1.28 -> -1.30 | No clear effect |
| replace zero-vector ablation with PAD-token ablation (ablation) | Jacobian Scopes · LAMBADA and IWSLT2017, six LLMs | best-method ranking | No clear effect |
| raise Fisher rank from r=1 to r=256 (ablation) | Fisher Scope · LAMBADA, LLaMA-3.2 1B | AOPC | No clear effect |
| keep only rank-1 principal Fisher direction (ablation) | Fisher Scope · LAMBADA, LLaMA-3.2 1B | AOPC | Improved |
| use Fisher Scope instead of Input x Gradient | Gemma-3 4B · LAMBADA | AOPC: -1.70 -> -1.81 | Improved |
Closest papers by meaning
- LLM Processes: Numerical Predictive Distributions Conditioned on Natural Language (NeurIPS 2024 · similarity 0.90)
- Scalable Influence and Fact Tracing for Large Language Model Pretraining (ICLR 2025 · similarity 0.87)
- DATE-LM: Benchmarking Data Attribution Evaluation for Large Language Models (NeurIPS 2025 · similarity 0.87)
- AttriBoT: A Bag of Tricks for Efficiently Approximating Leave-One-Out Context Attribution (ICLR 2025 · similarity 0.87)
- Consolidating Rewarded Perturbations for LLM Post-Training (NeurIPS 2026 · similarity 0.87)
- Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
