LLM Token Attribution

Hands-on with Captum

Wednesday, December 9, 2026 · 13:00–16:00
Online · in English

by Israel A. Azime and Alexander Blatt

We cordially invite you to our LLM Token Attribution workshop, organized by Saarland University. In this workshop, we will focus on hands-on exercises that demonstrate how to interpret the outputs of large language models (LLMs) using token attribution with Captum, the open-source interpretability library for PyTorch. The session is designed to be practical and interactive, allowing participants to explore model behavior through guided examples.

Who should attend?

Anyone who wants to understand why language models produce particular outputs. Participants will learn how to measure which input tokens contribute most to a model's generated text and gain insights into model behavior. Basic Python and PyTorch knowledge is recommended. Familiarity with what LLMs are is helpful; no prior interpretability experience is needed. The exercises run in Google Colab, so you only need a Google account and a browser.

What you will learn

Why token attribution? How attribution methods help explain which parts of a prompt drive an LLM’s output, and what the scores can and cannot tell you.

Getting started with Captum. Setting up Captum’s LLM attribution tools for a Hugging Face language model in Google Colab.

Perturbation-based attribution. Hands-on token attribution with feature ablation and Shapley values, including how the choice of baseline affects the results.

Gradient-based attribution. Hands-on token attribution using methods such as layer integrated gradients.

Interpreting the results. Visualizing attribution scores, comparing methods, and drawing sound conclusions about model behavior.

Format

The workshop is held online and consists of a theoretical introduction and practical sessions. These exercises will guide participants step-by-step through computing and interpreting token attributions for LLM outputs. Exercises include explaining a model’s free-text completions, analyzing a loan-approval prompt based on the German Credit dataset, and measuring how few-shot examples influence a sentiment classification.

Materials

Materials will be made available for registered participants before the workshop.

Registration

Eligibility	Members of academic institutions
Registration fee	None
Organization	CSHPC, Saarland University, as part of NHR Süd-West

The organization and coordination of this event are carried out by NHR Süd-West, a member of the National High Performance Computing (NHR) initiative in Germany. For more information, visit https://nhrsw.de/.