Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TAXI: Evaluating Categorical Knowledge Editing for Language Models

Derek Powell, Walter Gerych, Thomas Hartvigsen

TL;DR

TAXI introduces a taxonomy-based benchmark to evaluate consistency in categorical knowledge editing for LLMs, addressing the problem that edits should propagate to inherited properties to maintain a coherent world model. It defines categories $c \\in \\mathcal{C}$ with property sets $p^c$, subjects $s$, and a formal editing process that tests whether changing a subject’s category leads to correct changes in its properties through metrics including Edit Success, Invariance, and Consistency. Experiments on Llama-2 editors reveal high forward edit success but substantially lower property-consistency, with invariance outperforming consistency and atypical subjects editing more easily, indicating room for improved generalization. A human study shows near-perfect property generalization, underscoring the gap between current editors and human cognition and highlighting TAXI as a principled, human-solvable benchmark for advancing consistent model editing and safer factual updates.

Abstract

Humans rarely learn one fact in isolation. Instead, learning a new fact induces knowledge of other facts about the world. For example, in learning a korat is a type of cat, you also infer it is a mammal and has claws, ensuring your model of the world is consistent. Knowledge editing aims to inject new facts into language models to improve their factuality, but current benchmarks fail to evaluate consistency, which is critical to ensure efficient, accurate, and generalizable edits. We manually create TAXI, a new benchmark dataset specifically created to evaluate consistency in categorical knowledge edits. TAXI contains 11,120 multiple-choice queries for 976 edits spanning 41 categories (e.g., Dogs), 164 subjects (e.g., Labrador), and 183 properties (e.g., is a mammal). We then use TAXI to evaluate popular editors' categorical consistency, measuring how often editing a subject's category appropriately edits its properties. We find that 1) the editors achieve marginal, yet non-random consistency, 2) their consistency far underperforms human baselines, and 3) consistency is more achievable when editing atypical subjects Our code and data are available at https://github.com/derekpowell/taxi.

TAXI: Evaluating Categorical Knowledge Editing for Language Models

TL;DR

TAXI introduces a taxonomy-based benchmark to evaluate consistency in categorical knowledge editing for LLMs, addressing the problem that edits should propagate to inherited properties to maintain a coherent world model. It defines categories with property sets , subjects , and a formal editing process that tests whether changing a subject’s category leads to correct changes in its properties through metrics including Edit Success, Invariance, and Consistency. Experiments on Llama-2 editors reveal high forward edit success but substantially lower property-consistency, with invariance outperforming consistency and atypical subjects editing more easily, indicating room for improved generalization. A human study shows near-perfect property generalization, underscoring the gap between current editors and human cognition and highlighting TAXI as a principled, human-solvable benchmark for advancing consistent model editing and safer factual updates.

Abstract

Humans rarely learn one fact in isolation. Instead, learning a new fact induces knowledge of other facts about the world. For example, in learning a korat is a type of cat, you also infer it is a mammal and has claws, ensuring your model of the world is consistent. Knowledge editing aims to inject new facts into language models to improve their factuality, but current benchmarks fail to evaluate consistency, which is critical to ensure efficient, accurate, and generalizable edits. We manually create TAXI, a new benchmark dataset specifically created to evaluate consistency in categorical knowledge edits. TAXI contains 11,120 multiple-choice queries for 976 edits spanning 41 categories (e.g., Dogs), 164 subjects (e.g., Labrador), and 183 properties (e.g., is a mammal). We then use TAXI to evaluate popular editors' categorical consistency, measuring how often editing a subject's category appropriately edits its properties. We find that 1) the editors achieve marginal, yet non-random consistency, 2) their consistency far underperforms human baselines, and 3) consistency is more achievable when editing atypical subjects Our code and data are available at https://github.com/derekpowell/taxi.
Paper Structure (20 sections, 4 figures, 5 tables)

This paper contains 20 sections, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Methods
  4. Categories
  5. Properties
  6. Subjects
  7. Data Collection
  8. The taxi Dataset
  9. Metrics
  10. Experiments
  11. Edit success does not imply property success
  12. Edits exhibit greater property invariance than consistency
  13. Atypical subjects are easier to edit than typical subjects
  14. Consistency is consistent across superordinate categories
  15. Human Study
  16. ...and 5 more sections

Figures (4)

  • Figure 1: Consistent categorical edits reclassify subjects, which inherit properties of their new categories.
  • Figure 2: Consistency for forward multiple-choice test queries by editor and category (dashed line indicates chance).
  • Figure 3: Edit success for forward and reverse multiple-choice test queries by editor type. Dashed line indicates chance performance for forward and reverse edits by color.
  • Figure 4: Screenshot illustrating a trial of the of human annotation study.