Source-linked AI summary
Evaluating the Ripple Effects of Knowledge Editing in Language Models
Roi Cohen, Eden Biran, Ori Yoran, Amir Globerson, Mor Geva
TL;DR
Language-model knowledge editing is commonly evaluated on the edited fact and unrelated-fact preservation, leaving logically implied ripple effects underexamined. The paper proposes ripple-effect criteria and builds the 5K-entry RippleEdits benchmark; current methods perform poorly on these effects, while simple in-context editing scores best. The benchmark is diagnostic rather than exhaustive, and its data depend on potentially incomplete or outdated Wikidata.
Problem
Existing knowledge-editing evaluation focuses on the edited fact and unrelated-fact preservation, but does not fully test logically derived changes to related facts.
Method
The paper proposes six ripple-effect evaluation criteria and constructs RippleEdits, a benchmark of 5K factual edits with related test queries.
Results
Current knowledge-editing methods often fail to introduce consistent ripple-effect changes, while a simple in-context editing baseline achieves the best results on RippleEdits.
Takeaways & Limitations
Ripple effects should be considered in knowledge-editing evaluation and method development, with in-context editing identified as a promising direction.
Takeaways & Limitations
RippleEdits covers only a small fraction of possible ripple edits, and its data-generation pipeline relies on potentially incomplete or outdated Wikidata information.
Abstract
from arXiv · showhide
Modern language models capture a large body of factual knowledge. However, some facts can be incorrectly induced or become obsolete over time, resulting in factually incorrect generations. This has led to the development of various editing methods that allow updating facts encoded by the model. Evaluation of these methods has primarily focused on testing whether an individual fact has been successfully injected, and if similar predictions for other subjects have not changed. Here we argue that such evaluation is limited, since injecting one fact (e.g. ``Jack Depp is the son of Johnny Depp'') introduces a ``ripple effect'' in the form of additional facts that the model needs to update (e.g.``Jack Depp is the sibling of Lily-Rose Depp''). To address this issue, we propose a novel set of evaluation criteria that consider the implications of an edit on related facts. Using these criteria, we then construct RippleEdits, a diagnostic benchmark of 5K factual edits, capturing a variety of types of ripple effects. We evaluate prominent editing methods on RippleEdits, showing that current methods fail to introduce consistent changes in the model's knowledge. In addition, we find that a simple in-context editing baseline obtains the best scores on our benchmark, suggesting a promising research direction for model editing.
1 Introduction
Knowledge editing updates incorrect or outdated facts in language models, but standard evaluation largely checks only the edited fact and unrelated-fact preservation. This work evaluates the logically implied ripple effects of edits, introduces RippleEdits, and finds current parametric methods inconsistent while in-context editing performs best.
- Knowledge editing addresses factual beliefs that may be incorrect or outdated and affect downstream performance, reliability, and usability.
- Standard evaluation checks whether the edited fact is correctly generated while preserving unrelated facts, but does not test logically derived changes.
- Ripple effects are changes to other facts implied by a factual edit, such as sibling relations following a changed parent relation.
- The proposed criteria test integration with existing knowledge through logical reasoning, complex composition, subject aliasing, and relation specificity.
- RippleEdits contains 5K factual edits with test queries and metadata covering edit timestamps and entity popularity.
- Current knowledge-editing methods often fail on ripple effects, whereas a simple in-context editing baseline achieves the best benchmark results.
2 Problem Setting
The problem setting represents factual knowledge as subject-relation-object triplets and distinguishes modifying encoded facts from injecting new ones. The edit type depends on whether the information was already captured by the model.
- Factual knowledge is represented as triplets (e, r, o) containing a subject entity, relation, and object.
- A modification changes the object of an already encoded fact, whereas an injection adds a fact not captured by the model.
- For one-to-one relations, injection populates an empty object; for one-to-many relations, it augments the existing object set.
- Whether an edit is called modification or injection depends on whether the information was encoded before editing.
3 Ripple Effects of Factual Edits
A factual edit can require related facts within the model’s knowledge to change, remain unchanged, or be added. The evaluation focuses on ripple effects within two hops using logical, compositional, aliasing, and preservation criteria.
- The ripple effect R(δ) is the set of triplets the model needs to inject, modify, or delete to reflect the world after edit δ.
- Ripple severity is the number of facts affected by one edit, with examples ranging from many changes for Rome’s country to local sibling updates.
- The evaluation restricts ripple effects to modified facts within a 2-hop distance from the edit’s subject or object.
- Logical Generalization: Logical Generalization tests whether relation constraints such as symmetry and transitivity still hold after editing.
- Compositionality: Compositionality tests whether the model composes the edited relation with facts about the target object or another subject.
- Subject Aliasing and Preservation: Subject Aliasing requires equivalent edits for aliases, while Preservation requires existing objects in one-to-many relations to remain unchanged.
6. Relation Specificity (RS): We test that facts
Relation Specificity checks that editing one subject-relation pair does not alter the subject’s other, unrelated relations. This criterion complements Logical Generalization by testing preservation outside the edit’s dependency set.
- 6. Relation Specificity (RS): Relation Specificity verifies that other relations for the edited subject remain unchanged.For example, modifying Prince’s sibling should not change his mother.
- 6. Relation Specificity (RS): The criterion uses facts whose relations are not expected to depend on the edited relation.These facts are distinct from those tested by Logical Generalization.
4 The RIPPLEEDITS Benchmark
RIPPLEEDITS is generated from Wikidata facts through edit construction, ripple-effect query generation, and natural-language templating. The resulting benchmark contains diverse subsets and criteria-focused queries, with validation showing high semantic and grammatical quality.
- 4.1 Data Generation Pipeline: RIPPLEEDITS uses Wikidata triplets and manually defined relation-dependent rules to generate factual edits and ripple-effect test queries.The pipeline restricts edits and queries to 54 basic relations with manually crafted templates and logical rules.
- 4.1 Data Generation Pipeline: The benchmark includes RECENT, RANDOM, and POPULAR triplets to represent new facts, random correction edits, and edits involving popular entities.RECENT facts support injection edits, while RANDOM and POPULAR facts support modification edits.
- 4.1 Data Generation Pipeline: Test queries cover Logical Generalization, Relation Specificity, Compositionality I and II, Subject Aliasing, and Preservation.Queries test changed or retained facts after an edit, including two-hop compositions, aliases, and original relation targets.
- 4.2 Data Statistics: 18–26 test queries per edit and over 3 queries per evaluation test were generated on average, with diverse relation frequencies across subsets.POPULAR edits contain more popular subjects, while RECENT edits have more popular objects.
- 4.3 Data Quality: 100% of queries were judged semantically sound and 98.5% grammatically correct.The grammatically incorrect cases involved entity representations in a non-English language.
5 Experiments
The experiments evaluate RippleEdits using several knowledge-editing methods, language models, and an in-context baseline. Existing methods struggle to propagate edits consistently, while performance varies by model size, entity popularity, evaluation criterion, and editing method.
- 5.1 Evaluation Setting: The evaluation retains only edits that successfully apply and test queries the model answered correctly before editing.Each test generation allows up to 20 tokens and counts an answer as correct when it contains an alias of a target object.
- 5.2 Results: 38−66 average accuracy across models indicates that existing editing methods struggle with ripple effects despite making local factual updates.The results describe these changes as largely surface-level, without propagation to related facts.
- 5.2 Results: ICE achieves the best overall results, outperforming ROME by more than 10 points for GPT-Neo and 29 points for LLaMA across subsets.GPT-3 with ICE performs best on average, while 7B LLaMA is competitive on the RECENT and POPULAR subsets.
- 5.2 Results: Performance improves with model size, but ICE makes 7B LLaMA competitive with much larger GPT-3, so scaling alone may not fix current editing drawbacks.Across methods, MEND performs better on Logical Generalization, whereas ROME and MEMIT perform better on Compositionality I and II.
- 5.3 Error Analysis: At least 68% of failures across settings involve erroneous knowledge changes rather than no change; ROME introduces more incorrect changes, while ICE causes abstentions.Subject Aliasing is consistently high at ≥86.8, whereas other criteria vary substantially across models, methods, and edits.
6 Conclusion and Discussion
The paper introduces ripple effects as a missing dimension in knowledge-editing evaluation and develops criteria and a benchmark to assess them. It finds that prominent methods often produce inconsistent related updates, while identifying important benchmark scope limitations.
- Ripple effects capture the further updates to related facts implied by editing one fact.The paper frames these effects as a central concept for evaluating knowledge editing.
- The proposed evaluation criteria test whether related facts are updated or retained consistently after an edit.The criteria target logical and relational consequences of edits.
- RIPPLEEDITS is a diagnostic benchmark designed to evaluate how knowledge-editing methods handle ripple effects.It operationalizes the paper’s proposed evaluation focus.
- Prominent knowledge-editing methods often fail to introduce consistent edits that capture an edit’s ripple effects.The authors present this as evidence that current methods do not adequately handle related factual changes.
- The benchmark covers only a small fraction of possible ripple edits, including limited-hop and single-edit cases.Future work could examine longer chains, graph structure, and batches of multiple edits.
- RIPPLEEDITS relies on potentially incomplete or outdated Wikidata and omits some existing edit and distant-fact retention tests.The authors also note that automatically determining which facts should change remains difficult.