Research map / Large language models / Language, safety and interpretability
LLM Attributor: Interactive Visual Attribution for LLM Generation
Read the paperCheck your idea against 108,967 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.
Presents LLM Attributor, a Python library with interactive visualizations for training data attribution of LLM-generated text, built on an improved DataInf with checkpoint-median scoring and a side-by-side comparison view.
Problem
Training data attribution methods for LLMs exist algorithmically, but there is little work on how to present attribution results to people, and existing DataInf degrades on free-form prompts.
Approach
Wraps a DataInf-based attribution (extended with multiple shuffled checkpoints and median aggregation) in an interactive notebook-friendly library with a main attribution view and a comparison view for LLM-generated versus user-edited text.
Key results
- Demonstrates via two usage scenarios (Hawaii wildfire misinformation and finance QA with LLaMA2-13B-Chat) how attribution helps developers trace problematic generations to specific training data points.
- Library is installable via PyPI (pip install llm-attributor) and supports Jupyter, Colab and VSCode notebooks.
Limitations
- No quantitative evaluation of attribution accuracy or user study; validation is through illustrative usage scenarios only.
- DataInf shows limited performance on free-form prompts, which the authors address heuristically.
Experimental setup
Datasets: online articles about 2023 disasters wealth-alpaca-lora
Baselines: TracIn
Builds on: DataInf TracIn
Closest papers by meaning
- Scalable Influence and Fact Tracing for Large Language Model Pretraining (ICLR 2025 · similarity 0.87)
- Controlled Text Generation via Language Model Arithmetic (ICLR 2024 · similarity 0.86)
- Evaluating Large Language Models at Evaluating Instruction Following (ICLR 2024 · similarity 0.86)
- AttriBoT: A Bag of Tricks for Efficiently Approximating Leave-One-Out Context Attribution (ICLR 2025 · similarity 0.86)
- DATE-LM: Benchmarking Data Attribution Evaluation for Large Language Models (NeurIPS 2025 · similarity 0.86)
- LLMEval: A Preliminary Study on How to Evaluate Large Language Models (AAAI 2024 · similarity 0.86)
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.