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.
Introduction
Large Language Models (LLMs) often contain outdated, biased, or incorrect information. While researchers have developed "knowledge editing" techniques to surgically update specific facts without retraining, these methods often treat facts as isolated propositions. In reality, human knowledge is a dense web of interconnected facts. If you tell a human that a "Korat" is a type of cat, they immediately infer it has whiskers, claws, and fur.
This paper introduces TAXI (TAXonomic Inference), a benchmark designed to measure categorical consistency: the ability of a model to propagate an edit through a taxonomic hierarchy. Without this "ripple effect," editing a model becomes a game of Whac-A-Mole—you might change one label, but the model’s internal world model remains fractured and inconsistent.
The TAXI Benchmark
TAXI is a manually curated benchmark consisting of 976 edits across 41 categories (e.g., Dogs, Instruments, Vehicles) and 11,120 multiple-choice queries. A key feature of TAXI is the inclusion of typicality markers. Each category includes "typical" subjects (e.g., Labrador for dogs) and "atypical" subjects (e.g., Pekingese) to determine if the strength of prior knowledge affects the model's ability to accept and propagate counterfactual edits.
Methods
We evaluated three prominent editing approaches on the Llama-2 7B model:
- Fine-tuning (FT): Directly updating model weights on the new fact.
- ROME (Rank-One Model Editing): A surgical update to specific MLP layers treated as key-value memories.
- ICE (In-Context Editing): Providing the new fact as a prompt prefix.
Performance was evaluated using three metrics: Edit Success (did the model accept the new category?), Consistency (did it update properties that should change?), and Invariance (did it keep properties that should not change?). We also established a human baseline with 19 participants.
Key Results
Our findings reveal a massive gulf between a model's ability to repeat a new fact and its ability to understand the implications of that fact. While ICE achieved a perfect 1.0 success rate in changing labels, its consistency score was only 0.47, trailing far behind the human baseline of 0.86.
Key Insights
- The Consistency Gap: Models are "stubborn" reasoners. They are much better at maintaining invariance (~90%) than they are at updating properties that should change (~45%).
- The Typicality Effect: It is significantly easier to consistently edit atypical subjects. Strong prior associations for typical subjects in the training data (e.g., a Labrador's strong link to "dog") make counterfactual edits harder to propagate.
- Directionality Issues: Current editors fail almost entirely on reverse queries (e.g., "Name a type of dog" resulting in "cobra"). This suggests these methods patch specific input-output mappings rather than updating the model's underlying knowledge representation.
Conclusion
TAXI demonstrates that current knowledge editing methods achieve only marginal consistency. The failure to propagate edits through taxonomic hierarchies suggests that LLMs still treat facts as flat lists rather than structured mental models. Future research must move beyond superficial label patching toward methods that update the model's internal logic holistically.
Summary updated