PaperlayerView exact revision on arXiv

Source-linked AI summary

MQuAKE: Assessing Knowledge Editing in Language Models via Multi-Hop Questions

Zexuan Zhong, Zhengxuan Wu, Christopher D. Manning, Christopher Potts, Danqi Chen

arXiv:2305.14795v3cs.CL

TL;DR

Knowledge-editing evaluations have largely tested direct fact recall, leaving unclear whether edits propagate to entailed multi-hop answers. The paper introduces MQUAKE and MeLLo, finding that existing editors fail on such questions while MeLLo substantially outperforms them and scales to large language models.

  • Problem

    Existing evaluations mainly test recall of edited facts, leaving unanswered whether models update answers that change as entailed consequences of those edits.

  • Method

    MQUAKE benchmarks multi-hop consequences using counterfactual and temporal datasets, while MeLLo stores edits externally and iteratively prompts the model to answer consistently with them.

  • Results

    Existing editors often recall edited facts but fail drastically on consequential multi-hop questions; MeLLo significantly outperforms them, including at large edit counts.

  • Takeaways & Limitations

    MQUAKE provides a more complete evaluation of knowledge editing, and MeLLo offers a lightweight approach applicable to large LMs without additional training.

  • Takeaways & Limitations

    The evaluation mainly covers GPT-J and Vicuna, while MeLLo requires manually defined prompts for new tasks.

Abstract

from arXiv · show

The information stored in large language models (LLMs) falls out of date quickly, and retraining from scratch is often not an option. This has recently given rise to a range of techniques for injecting new facts through updating model weights. Current evaluation paradigms are extremely limited, mainly validating the recall of edited facts, but changing one fact should cause rippling changes to the model's related beliefs. If we edit the UK Prime Minister to now be Rishi Sunak, then we should get a different answer to Who is married to the British Prime Minister? In this work, we present a benchmark, MQuAKE (Multi-hop Question Answering for Knowledge Editing), comprising multi-hop questions that assess whether edited models correctly answer questions where the answer should change as an entailed consequence of edited facts. While we find that current knowledge-editing approaches can recall edited facts accurately, they fail catastrophically on the constructed multi-hop questions. We thus propose a simple memory-based approach, MeLLo, which stores all edited facts externally while prompting the language model iteratively to generate answers that are consistent with the edited facts. While MQuAKE remains challenging, we show that MeLLo scales well with LLMs (e.g., OpenAI GPT-3.5-turbo) and outperforms previous model editors by a large margin.

1 Introduction

MQUAKE addresses whether knowledge-editing methods can update entailed consequences, not merely recall directly edited facts. Existing methods often fail this multi-hop evaluation, while MeLLo uses external memory and iterative prompting to improve consistency.

  • LLMs need up-to-date knowledge without the massive retraining costs associated with rebuilding models.
  • Existing benchmarks mainly test recall of edited facts and preservation of unrelated knowledge, leaving consequential answer changes unaddressed.
  • MQUAKE evaluates 2-, 3-, and 4-hop questions requiring edited models to propagate changes through chains of entailed facts.
  • GPT-J edited by ROME answered 7.4% of MQUAKE-CF multi-hop questions after editing, versus 40.5% before editing.
  • MeLLo stores edits externally, decomposes questions into successive sub-questions, checks consistency, and significantly outperforms existing editors without additional training.

2 Problem Definition

The paper defines knowledge editing as changing stored facts and evaluates success by whether those changes support correct multi-hop consequences. Its problem definition contrasts this with prior single-hop recall tests.

  • A fact edit replaces an object in a subject-relation-object triple with a new object, written e = (s,r,o → o*).
  • Knowledge editing maps a language model and edit collection to an edited model, either by modifying weights or using edits as external guidance.
  • Prior evaluations test whether edited models recall updated facts under paraphrased prompts and preserve unrelated knowledge.
  • A multi-hop chain links each fact’s object to the next fact’s subject, so a question about the first subject targets the final tail entity.
  • After edits change one or more chain facts, the model must use updated knowledge to answer the resulting entailed question.

3 MQUAKE: Multi-hop Question Answering for Knowledge Editing

MQUAKE is a Wikidata-based benchmark with counterfactual and temporal datasets for testing multi-hop knowledge editing. Its instances encode edits, questions, pre- and post-edit answers, and fact chains, with multiple metrics for evaluation.

  • MQUAKE contains MQUAKE-CF for counterfactual edits and MQUAKE-T for temporal updates to outdated real-world knowledge.
  • Data construction: The benchmark samples coherent 2-, 3-, and 4-fact chains from a Wikidata subgraph covering 37 common relations and common entities.
  • Data construction: Chains containing facts GPT-J cannot recall are filtered out before multi-hop questions are generated with ChatGPT.
  • Data construction: MQUAKE-CF samples one or more counterfactual edits, whereas MQUAKE-T uses authentic changes between April 2021 and April 2023 Wikidata dumps and one edit per instance.
  • Dataset format: Each dataset instance contains edits, three questions, pre- and post-edit answers, and corresponding pre- and post-edit fact chains.
  • Dataset statistics: MQUAKE-CF has more than 9K N-hop questions, with 3K used for experiments, while MQUAKE-T contains 1.8K instances.
  • Evaluation metrics: Instance-wise accuracy measures whether all individual single-hop facts in a multi-hop instance are recalled, before and after editing.
  • Evaluation metrics: Multi-hop accuracy is the main metric and counts an instance as accurate when any of its three generated questions is answered correctly.

4 MQUAKE Challenges Model Editors

MQUAKE evaluation shows that knowledge editors can recall edited facts yet fail to propagate those edits through multi-hop reasoning, especially as edit volume grows. MeLLo instead keeps edits in explicit memory and iteratively checks generated answers against them.

  • Experimental setup: GPT-J and Vicuna-7B evaluate fine-tuning, MEND, ROME, and MEMIT using edit-wise, instance-wise, and multi-hop accuracy.The setup includes cloze-based editing, in-context demonstrations, and chain-of-thought prompting for multi-hop evaluation.
  • Single-instance results: MEMIT and ROME exceed 90% edit-wise accuracy with GPT-J and Vicuna-7B when injecting small numbers of edits.These results indicate strong recall of the injected facts at inference time.
  • Multi-hop results: 40.5% → 7.0% with GPT-J and 30.2% → 4.9% with Vicuna-7B: MEMIT multi-hop performance collapses after editing.All evaluated editors fail on questions whose answers should change as consequences of edited facts, and performance also drops relative to the pre-edit models.
  • MeLLo: MeLLo freezes the base model, stores edits explicitly, decomposes questions into subquestions, and self-checks tentative answers against retrieved edited facts.The method is designed to support large black-box language models while enforcing consistency with external edits.
  • Multi-hop results: Chain-of-thought prompting improves results slightly across settings but still leaves multi-hop performance catastrophically poor.This supports the conclusion that the problem is not resolved by prompt-based reasoning alone.
  • Temporal updates: On MQUAKE-T, all methods except fine-tuning achieve near-perfect edit-wise and instance-wise accuracy, yet multi-hop performance falls significantly below the base model.MEND performs comparatively well with CoT in this setting, possibly because it edits certain relations effectively.
  • Scaling edits: Injecting more edits simultaneously further reduces multi-hop performance on both MQUAKE-CF and MQUAKE-T.The multi-edit evaluation uses k ∈ {1, 100, 1000, 3000} for MQUAKE-CF and k ∈ {1, 100, 500, 1868} for MQUAKE-T.

5 MeLLo: A Proposal for Editing Large Language Models

MeLLo keeps the base language model frozen, stores edits explicitly, and iteratively decomposes and checks multi-hop answers for consistency. It outperforms existing editors across settings and requires no training, but depends on retrieving all associated edited facts.

  • Evaluation Results: Existing knowledge-editing methods fail catastrophically on MQUAKE multi-hop questions.The evaluation motivates MeLLo as an alternative to weight-based editing.
  • Method: MeLLo keeps the base language model frozen and stores edited facts in an explicit external memory.Edited facts are converted into statements and indexed with a pretrained Contriever retrieval model.
  • Method: MeLLo decomposes multi-hop questions into subquestions, generates tentative answers, and self-checks them against edited facts before returning a final answer.The procedure uses the model's unedited knowledge to generate tentative answers, then checks consistency with retrieved edits.
  • Evaluation Results: MeLLo significantly outperforms MEMIT and MEND across settings with GPT-J, and the performance gap grows substantially with GPT-3.5.The method is also described as more efficient and requiring no training.
  • Evaluation Results: Correct multi-hop answering requires MeLLo's retriever to retrieve all associated edited facts from memory.The paper investigates the effect of retrieval accuracy using GPT-3.5.

6 Related Work

Related work includes weight-based and auxiliary-model approaches to knowledge editing, evaluation of factual updates and side effects, and prompting-based multi-hop QA. MeLLo differs by using the base model itself to self-check retrieved edits, enabling black-box use without extra training.

  • Knowledge-editing methods: Prior knowledge-editing methods inject new knowledge into static model artifacts, including by modifying concept-related weights or using auxiliary components.These approaches target incremental updates and editing at scale.
  • Prompting methods for multi-hop QA: Unlike SERAC, MeLLo directly uses the base model to self-check generations, making it applicable to black-box language models without extra training.This design avoids training additional models to incorporate memory.
  • Knowledge-editing evaluation: Existing evaluation commonly tests recall of edited or related facts and whether irrelevant knowledge remains unchanged.More recent work also studies inference from injected facts and unintended side effects.
  • Knowledge-editing evaluation: MQUAKE complements prior evaluation by testing multi-hop answers that should change as entailed consequences of edits.The benchmark specifically exposes failures on these questions.
  • Prompting methods for multi-hop QA: Prompting methods combine language models with optional retrieval or external NLI modules for multi-step question answering.MeLLo instead relies on the language model itself to perform step-by-step NLI.

7 Conclusion

The paper introduces MQUAKE to evaluate whether edits propagate to entailed multi-hop consequences and proposes MeLLo as a training-free alternative. Edited models recall edited facts but fail on these consequences, while MeLLo significantly outperforms existing methods.

  • Conclusion: MQUAKE evaluates knowledge editing through multi-hop questions whose answers should change as consequences of edited facts.The benchmark is intended to support research on faithful knowledge editing.
  • Conclusion: Edited language models effectively recall edited facts but fail on multi-hop questions that are entailed consequences of those edits.This contrasts direct fact recall with propagated knowledge changes.
  • Conclusion: MeLLo is a simple alternative that significantly outperforms existing knowledge-editing methods without additional training and can be applied to large language models such as GPT-3.5.The conclusion presents MeLLo as a direction for developing faithful knowledge-editing methods.

Limitations

The paper's limitations concern evaluation scope, model scale, prompt dependence, and question authorship. These boundaries leave broader model coverage, smaller-model performance, new-task prompting, and human-authored questions for future work.

  • Evaluation scope: Existing knowledge-editing methods are mainly evaluated on GPT-J and Vicuna, leaving their efficacy on other LLMs less explored.The paper notes that these methods are computationally expensive and defers broader evaluation.
  • Model scale: MeLLo's demonstrated advantage is limited to models with more than 6B parameters, while its performance on smaller models such as GPT-2 remains future work.MeLLo relies on language models for question decomposition and self-checking.
  • Task adaptation: MeLLo requires manually defined prompts to drive language models on new tasks.The paper leaves evaluation across other tasks for future work.
  • Question authorship: MQUAKE's multi-hop questions are automatically generated by ChatGPT rather than crafted by humans.The authors suggest human-authored questions could better align the benchmark with realistic applications.

A Details of Dataset Construction

The dataset construction samples coherent Wikidata fact chains, filters for recallability, and uses ChatGPT to generate logically consistent multi-hop questions. Experiments use a 3,000-instance conflict-free subset of MQUAKE-CF.

  • Sampling fact chains: Fact chains contain N = {2,3,4} triples sampled from Wikidata under heuristics designed to ensure coherence and natural questions.The constraints exclude circles and repeated relation types, limit object types, restrict country objects to later hops, and arrange person or location objects consecutively.
  • Filtering and recall: Chains containing at least one fact GPT-J cannot recall are filtered out, and each relation receives a question template with eight demonstrations.In-context learning is used to capture the desired answer format.
  • Question generation: ChatGPT converts each fact chain into multi-hop questions using a dedicated prompt.The generated questions are reported as grammatically correct and logically consistent with the triple-chain flow.
  • Evaluation subset: Experiments use a 3,000-instance subset of MQUAKE-CF drawn from the full dataset.The subset is used for efficient evaluation and avoids knowledge conflicts across instances.

C Evaluation Metrics

The paper evaluates edited models through edit recall, instance-level consistency, and multi-hop question answering. These metrics distinguish direct recovery of edited facts from complete instance and compositional performance.

  • Metric overview: The evaluation uses edit-wise success rate, instance-wise accuracy, and multi-hop accuracy for an edited model f∗(⋅).The metrics respectively target edited-fact recall, complete fact recall within instances, and multi-hop question answering.
  • Edit-wise success rate: Edit-wise success rate averages whether the edited model recalls each updated fact as its edited object.For edit e = (s,r,o → o∗), success is 1[f∗(tr(s)) = o∗].
  • Instance-wise accuracy: Instance-wise accuracy measures whether all facts associated with an instance are recalled before or after editing.The metric is defined separately for the original and edited model settings.
  • Multi-hop accuracy: Multi-hop accuracy counts an instance as correct when any associated multi-hop question is answered correctly.The metric is defined both before and after editing.

D Implementation Details for Knowledge Editing Methods

The experiments compare several weight-editing baselines and prompting configurations on MQUAKE, including chain-of-thought and the MeLLo retrieval-and-self-checking procedure. The MQUAKE-T setup broadens allowable labels for outdated answers.

  • Fine-tuning baseline: Fine-tuning updates selected GPT-J or Vicuna-7B parameters by gradient descent, using a weight-change norm constraint.The implementation fine-tunes layer 21 of GPT-J and layer 31 of Vicuna-7B with coefficient 5 × 10^-5.
  • Model editors: MEND and ROME are configured for GPT-J and Vicuna-7B using model-specific training, layers, hyperparameters, and covariance statistics.MEND uses a pretrained GPT-J editor and a separately trained Vicuna-7B editor; ROME updates GPT-J and Vicuna-7B at specified layers.
  • Model editors: MEMIT updates GPT-J or multiple Vicuna-7B layers with default hyperparameters and covariance statistics computed or released for each model.Vicuna-7B uses layers {5,6,7,8,9} and covariance statistics based on 100,000 Wikitext samples.
  • Prompting: Chain-of-thought prompting is used to maximize model performance on multi-hop questions.A simplified CoT prompt template is provided for the experiments.
  • Evaluation setup: MQUAKE-CF experiments use a conflict-free 3,000-instance subset, while retrieval accuracy is evaluated for groups of instances in MeLLo.The retrieval table defines retrieval accuracy as the fraction of instances for which all associated edited facts are correctly retrieved.
  • MQUAKE-T setup: For MQUAKE-T, the evaluation extends golden labels to include outdated answers because smaller language models tend to be less calibrated.GPT-3.5 is queried for outdated answers to construct the expanded labels.
  • MeLLo: MeLLo decomposes subquestions, retrieves relevant edited facts, and iteratively self-checks whether retrieved facts contradict generated answers.The procedure repeats until the model produces the final answer, with generated subquestions and retrieved facts serving distinct roles.

H Breakdown Results on MQUAKE-CF

Breakdown results on MQUAKE-CF show that multi-hop performance declines with longer questions and with more associated edits. MeLLo’s performance depends on retrieving all edited facts required by each question.

  • Hop count: 2-hop questions achieve much higher performance than 3-hop and 4-hop questions across editing methods using GPT-J.The breakdown is reported in Tables 11 and 12 under chain-of-thought prompting.
  • Number of edits: Performance worsens as the number of edits associated with an instance increases.This pattern is reported for questions involving 1, 2, 3, or 4 edits.
  • Retrieval dependence: MeLLo requires retrieval of all associated edited facts to answer a multi-hop question correctly.Each question is associated with 1–4 edited facts, motivating analysis of retrieval accuracy.
  • Retrieval dependence: More advanced retrieval techniques could improve MeLLo’s performance.The retrieval-accuracy analysis links retrieval quality to multi-hop performance without specifying a particular improved technique.
  • Templates: MQUAKE uses question templates for single-hop queries and cloze-style statement templates to convert edited facts into natural-language statements.These templates support GPT-J filtering and edited-fact representation.
Loading 2305.14795v3…