Paper deep dive
The Mirage of Model Editing: Revisiting Evaluation in the Wild
Wanli Yang, Fei Sun, Jiajun Tan, Xinyu Ma, Qi Cao, Dawei Yin, Huawei Shen, Xueqi Cheng
Models: LLMs evaluated with MEMIT, LLMs evaluated with ROME
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 94%
Last extracted: 3/12/2026, 6:17:40 PM
Summary
The paper 'The Mirage of Model Editing: Revisiting Evaluation in the WILD' critiques current model editing evaluation practices, identifying that synthetic benchmarks often use teacher forcing and ground-truth length truncation to artificially inflate performance. The authors introduce QAEdit, a real-world QA benchmark, and WILD, a task-agnostic evaluation framework, demonstrating that state-of-the-art editing methods perform significantly worse (38.5% vs 96.8%) under realistic conditions and fail to scale in sequential editing tasks.
Entities (5)
Relation Signals (3)
Teacher Forcing â causesoverestimationin â Model Editing Performance
confidence 95% · teacher forcing... leading to overestimated performance.
WILD â revealslimitationsof â Synthetic Evaluation
confidence 95% · switching from synthetic to WILDevaluation consistently leads to a significant performance decline
QAEdit â evaluates â Model editing
confidence 90% · we introduce QAEdit, a new benchmark... to rigorously assess model editing in real-world QA.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Despite near-perfect results reported in the literature, the effectiveness of model editing in real-world applications remains unclear. To bridge this gap, we introduce QAEdit, a new benchmark aligned with widely used question answering (QA) datasets, and WILD, a task-agnostic evaluation framework designed to better reflect real-world usage of model editing. Our single editing experiments show that current editing methods perform substantially worse than previously reported (38.5% vs. 96.8%). We demonstrate that it stems from issues in the synthetic evaluation practices of prior work. Among them, the most severe is the use of teacher forcing during testing, which leaks both content and length of the ground truth, leading to overestimated performance. Furthermore, we simulate practical deployment by sequential editing, revealing that current approaches fail drastically with only 1000 edits. This work calls for a shift in model editing research toward rigorous evaluation and the development of robust, scalable methods that can reliably update knowledge in LLMs for real-world use.
Tags
Links
Trouble viewing inline? Open PDF directly â
Full Text
71,417 characters extracted from source content.
Expand or collapse full text
arXiv:2502.11177v5 [cs.CL] 31 May 2025 The Mirage of Model Editing: Revisiting Evaluation in theWILD Are We Really Making Much Progress? Wanli Yang Ú Fei Sun Ăš# Jiajun Tan Ú Xinyu Ma t Qi Cao Ăš Dawei Yin t Huawei Shen Ú Xueqi Cheng Ú Ú State Key Laboratory of AI Safety, Institute of Computing Technology, CAS  University of Chinese Academy of Sciences t Baidu Inc. yangwanli24z@ict.ac.cn # sunfei@ict.ac.cn Abstract Despite near-perfect results reported in the liter- ature, the effectiveness of model editing in real- world applications remains unclear. To bridge this gap, we introduce QAEdit, a new bench- mark aligned with widely used question answer- ing (QA) datasets, andWILD, a task-agnostic evaluation framework designed to better reflect real-world usage of model editing. Our sin- gle editing experiments show that current edit- ing methods perform substantially worse than previously reported (38.5% vs. 96.8%). We demonstrate that it stems from issues in the synthetic evaluation practices of prior work. Among them, the most severe is the use of teacher forcingduring testing, which leaks both content and length of the ground truth, leading to overestimated performance. Furthermore, we simulate practical deployment by sequen- tial editing, revealing that current approaches fail drastically with only 1000 edits. This work calls for a shift in model editing research to- ward rigorous evaluation and the development of robust, scalable methods that can reliably up- date knowledge in LLMs for real-world use 1 . 1 Introduction âIf you canât measure it, you canât improve it.â â Lord Kelvin Model editing (Yao et al., 2023; Wang et al., 2024d) has attracted widespread attention for its promising vision: enabling efficient and precise updates to specific knowledge within pretrained Large Lan- guage Models (LLMs) without retraining from scratch. Recent advances report near-perfect re- sults on corresponding benchmarks (Meng et al., 2022; Wang et al., 2024b), suggesting substantial progress toward this goal. However, these results often come from synthetic, oversimplified evalu- ation settings (e.g., identical prompts for editing # Corresponding author: Fei Sun (sunfei@ict.ac.cn) 1 Code and data are released athttps://github.com/W anliYoung/Revisit-Editing-Evaluation. ZsRE COUNTERFACT QAEdit 40 60 80 100 ROME Success Rate (%) ZsRE COUNTERFACT QAEdit 0 20 40 60 80 100 WISE Success Rate (%) synthetic evaluationWILDevaluation Figure 1: Comparison of synthetic andWILDevaluation for ROME and WISE on Llama-2-7b-chat. and testing; more in §4) that may fail to capture real-world complexities. This disparity raises a crit- ical question:Can these promising results in the literature translate to practical applications? To address this question, we propose to study model editing in QA tasks, which provide clear evaluation criteria and broad applicability. This adaptation involves two key components: a real- world dataset and realistic evaluation. For dataset, we createQAEdit, a tailored dataset derived from three widely-used QA datasets, enabling editing methods to update LLMs with answers grounded in real-world tasks. For evaluation, we proposeWILD (WithoutIntervention,LiveDecoding), a task- agnostic evaluation framework that follows stan- dard QA evaluation protocols (Gao et al., 2024), assessing editing methods via the performance of edited LLMs on their previously failed questions. Our initial study reveals that current advanced editing methods achieve only a38.5% average suc- cess rate on QAEdit, significantly lower than the results reported in previous studies. This raises a question:Does the performance decline stem from QAEditâs real-world complexity, or from the shift of synthetic toWILDevaluation? To enable rigorous analysis, starting with single editing experiments, we evaluate six representative methods across three leading LLMs on QAEdit and two established editing benchmarks, using both evaluation frameworks. As illustrated in Figure 1, switching from synthetic toWILDevaluation con- sistently leads to a significant performance decline across editing methods and datasets. This drama- tic performance gap raises two critical questions: What differences between these frameworks drive such disparity, and which one most accurately re- flects editing effectiveness? To answer them, we carefully examine the setups for both synthetic andWILDevaluations. From this, we abstract four key modules (input,gener- ation strategy,output truncation, andmetric) and analyze their variations through controlled experi- ments. The results expose four critical limitations in current synthetic evaluation in model editing: â¶input module: using identical prompts for edit- ing and testing overlooks the variability and unpredictability in real-world queries; â· generation strategy: teacher forcing, which feeds the ground truth as input during decoding, artificially beautifies results by disregarding po- tential errors in the modelâs own outputs; âžoutput truncation: using target answer length to truncate outputs conceals errors (e.g., repe- tition, irrelevant, or incorrect information) that would occur with natural stopping criteria; âčmetric: match ratio may inflate performance by rewarding partial matches of incorrect answers. Among these issues,teacher forcingandtarget length truncationcause the most significant over- estimation, as they rely on ground truth that is un- available in real-world scenarios. This highlights thatsynthetic evaluation, reliant on such ideal- ized or even unrealistic conditions, fails to accu- rately measure true editing effectiveness. After uncovering evaluation issues via single editing analysis, we return to our initial question: how do editing methods perform under realistic conditions? In practice, editing requests arrive continuously, making sequential editing a more genuine test of real-world applicability. Under WILDevaluation, our sequential editing experi- ments show that current methods catastrophically fail to scale, with average success rates dropping to âŒ10% for only 1000 samples. Our work, for the first time, exposes severe is- sues in current evaluation of model editing research and demonstrates substantial limitations of existing editing methods under real-world conditions. We hope this work will inspire more rigorous evalua- tion practices and motivate the development of al- gorithms that can truly fulfill the promise of model editing: to reliably and scalably update knowledge in LLMsfor real-world applications. Our main contributions are as follows. âąWe introduce QAEdit, a benchmark tailored for real-world QA tasks, and establish a more rigorous evaluation framework, WILD. âąWe reveal that published model editing results are significantly inflated, and trace this overes- timation to issues in synthetic evaluation prac- tices, identified through modular analysis. âąWe expose the severe scalability challenges of current editing methods in practical applica- tions through sequential editing experiments. 2 Related Works 2.1 Model Editing Methodologies Existing model editing methods can be categorized into the following four types: Extension Based.These methods update LLMs by adding trainable parameters to encode new knowledge, e.g., additional neurons in FFN (Dong et al., 2022; Huang et al., 2023) or specialized mem- ory modules (Hartvigsen et al., 2023; Wang et al., 2024b), while preserving pretrained weights. Fine-tuning Based.Fine-tuning offers a straight- forward approach to update LLMsâ knowledge but faces catastrophic forgetting. Recent works miti- gate this by constraining parameter changes (Zhu et al., 2020) or leveraging Parameter-Efficient Fine- Tuning (PEFT) (Han et al., 2024) to limit modifica- tion scope (Yu et al., 2024; Wang et al., 2024a). Meta Learning.Employing meta learning, KE (De Cao et al., 2021), MEND (Mitchell et al., 2022), and MALMEN (Tan et al., 2024) train hyper- networks to predict effective gradients or parameter alterations for knowledge integration. Locate-Then-Edit.Based on the investigation of knowledge mechanisms in LLMs (Geva et al., 2021, 2022), KN (Dai et al., 2022), ROME (Meng et al., 2022), and PMET (Li et al., 2024b) utilize knowledge attribution and causal tracing to pin- point target knowledge to specific parameters, then perform localized editing. Furthermore, MEMIT (Meng et al., 2023) and EMMET (Gupta et al., 2024c) extend this for massive editing in a batch. 2.2 Evaluation of Model Editing Current evaluation of model editing primarily fo- cuses on editing effectiveness and side effects on model capabilities. Effectiveness of Editing.The effectiveness of editing is typically evaluated from four key prop- erties using artificial benchmarks and simplified evaluation settings:â¶reliability, success rate of editing;â·generalization, adaptability of edited knowledge to paraphrased prompts;âžlocality, im- pact on irrelevant knowledge;âčportability, appli- cability of edited knowledge in factual reasoning. For detailed information, We refer readers to Yao et al. (2023). In addition to these basic metrics, domain-specific editing tasks have been introduced, e.g., privacy preservation (Wu et al., 2023), bias mitigation (Chen et al., 2024b), and harm injection (Chen et al., 2024a). Side Effects of Editing.Recent research has also examined the potential side effects of editing on LLMs (Hoelscher-Obermaier et al., 2023; Li et al., 2024c). While locality shares similar objectives, its limited evaluation scope fails to capture the full extent of editing side effects. Recent studies (Yang et al., 2024a; Gu et al., 2024; Gupta et al., 2024b) have revealed that model editing can significantly compromise LLMsâ downstream tasks capabilities, motivating a growing research to mitigate such side effects (Ma et al., 2025; Fang et al., 2025). Discussion.In contrast to prior efforts that either benchmark editing algorithms on synthetic datasets or analyze their side effects, this work offers the first systematic re-examination of model editing un- der realistic deployment conditions. While AKEW (Wu et al., 2024) shares our motivation of advanc- ing model editing toward more realistic use cases, it pursues this goal by applying editing to a more complex task: unstructured editing. Our study in- stead re-evaluates the effectiveness of existing edit- ing methods on the same basic QA tasks adopted in prior work, but under a more rigorous and realistic evaluation protocol, revealing their limited practi- cal utility and uncovering the pitfalls of traditional editing evaluation. 3 QAEdit Motivation.While existing work reports remark- able success of model editing techniques (Meng et al., 2022; Wang et al., 2024b), their effective- ness in real-world applications remains unclear. To rigorously examine their practical utility, we focus on the most fundamental and widely studied task of QA rather than more complex settings such as multi-hop and unstructured editing. This choice is "Edit Prompt": "To whom was Grete Stern married?", "Edit Target": "Horacio Coppola", "Subject": "Grete Stern", "Rephrased Prompt": "Who was the spouse of Grete Stern?", "Locality Prompt": "When was the clock tower built in London?", "Locality Answer": "1859" Figure 2: An example from QAEdit. MethodFT-M MEND ROME MEMIT GRACE WISE Avg. Accuracy0.6110.3330.5850.5520.012 0.216 0.385 Table 1: Accuracy of edited Llama-2-7b-chat on ques- tions it failed before editing in QAEdit. motivated by a simple premise: if current editing methods struggle on basic QA tasks, then they are unlikely to succeed in more challenging scenarios, whereas failure in such tasks does not entail failure on the basic QA task. Specifically, we apply editing methods to correct LLMsâ errors in QA tasks and assess the improve- ment by re-evaluating edited LLMs on a standard QA evaluation framework, lm-evaluation-harness (Gao et al., 2024). Benchmark Preparation.Since existing edit- ing benchmarks are not derived from or aligned with mainstream QA tasks, we introduce QAEdit, a tailored benchmark to rigorously assess model editing in real-world QA. Specifically, QAEdit is constructed from three widely-used QA datasets with broad real-world coverage: Natural Questions (Kwiatkowski et al., 2019), TriviaQA (Joshi et al., 2017), and SimpleQA (Wei et al., 2024). Details about these datasets are provided in Appendix A.1. While these benchmarks provide questions and answers asedit promptsandtargetsrespectively, they lack essential fields that mainstream editing methods require for editing and evaluation. To ob- tain requiredsubjectsfor editing, we employ GPT-4 (gpt-4-1106-preview) to extract them directly from the questions. To align with the previous editing evaluation protocol, we evaluate: i)reliabilityus- ing original edit prompts; i)generalizationthrough GPT-4 paraphrased prompts; and i)localityusing unrelated QA pairs from ZsRE locality set 2 . As a result, QAEdit contains 19,249 samples across ten categories, ensuring diverse coverage of QA scenarios. Figure 2 shows a QAEdit entry with all fields. Dataset construction and dataset statistics are detailed in Appendix A.2. 2 We excludeportabilityevaluation as it concerns reasoning rather than our focus on knowledge updating in real-world. Who wrote the song âIf I Were a Boyâ ? âcontext-free input <BOS> BC Jean andToby âteacher forcing BeyoncĂ©JeanisTobyGad BCJeanandTobyGad Ground Truth: Output: STOP ground truth length â â â â â 3/5 match ratio â Edited LLM (a) synthetic evaluation framework ContextWho wrote the song âIf I Were a Boyâ ? âcontext-guided input <BOS> BeyoncĂ©isthewriter âautoregressive decoding BeyoncĂ©isthewriter<|endoftext|> BCBC Jean and Toby GadJeanandTobyGad Ground Truth: Output: STOP natural stopping criteria â Ă LLM-as- a-Judge â 0 Edited LLM (b) WILDevaluation framework Figure 3: Illustration of synthetic andWILDevaluation frameworks for measuring reliability, generalization, and locality. Each framework comprises four key modules:âinput,âgeneration strategy,âoutput truncation, and âmetric. Here, we use LLM-as-a-Judge as an example metric to illustrateWILD, which supports various metrics. Preliminary Study.We conduct single-edit ex- periments on Llama-2-7b-chatâs failed questions in QAEdit (detailed in §5). As shown in Table 1, after applying SOTA editing methods, the edited models achieve only 38.5% average accuracy under QA evaluation, far below previously reported results (Meng et al., 2023; Wang et al., 2024b). This raises a critical question:Is the performance degradation attributed to the real-world complexity of QAEdit, or to real-world QA evaluation? 4 A Tale of Two Evaluation Frameworks To identify the cause of this performance gap and guide further investigation, we first delve into the experimental setup of both editing (synthetic) and QA task (WILD) evaluations. We abstract them into four key modules:input,generation strategy,out- put truncation, andmetric. This modular paradigm enables systematic comparison between the two evaluation frameworks, as shown in Figure 3. Synthetic.We formalize the evaluation pipeline commonly used in prior model editing works (Yao et al., 2023; Wang et al., 2024b) assyntheticevalu- ation framework, which implements the four mod- ules in an idealized and overly simplified way (Fig- ure 3a):â¶input: using only question without addi- tional context;â·generation strategy: employing teacher forcing to feed ground truth tokens as input during decoding 3 ;âžoutput truncation: truncat- ing output to match the length of target answer; âčmetric: using token-level match ratio between the target and generated answer as accuracy. 3 The code snippets of mainstream editing evaluations with teacher forcing are presented in Appendix A.3. ModulesyntheticWILD Inputcontext-freecontext-guided Gen. Strategyteacher forcingautoregressive decoding Output Trunc.ground truth lengthnatural stopping criteria Metricmatch ratioLLM-as-a-Judge / EM Table 2: Key settings of synthetic andWILDevaluation across all four modules. WILD.We propose theWILD(WithoutInterven- tion,LiveDecoding) evaluation framework based on the standard QA evaluation protocol (Gao et al., 2024), which implements the core modules in a more realistic manner (Figure 3b):â¶input: pre- fixing question with contexts like task instructions; â·generation strategy: adopting autoregressive decoding, where each output serves as input for subsequent generation;âžoutput truncation: us- ing predefined stop tokens (e.g., â.â, â â, and â<|endoftext|>â) as signal to terminate genera- tion;âčmetric: WILDsupports evaluation metrics, including BERTScore (Zhang et al., 2020) and ex- act match (EM). Given its popularity and alignment with human judgment, we adopt LLM-as-a-Judge 4 (Li et al., 2024a) as the primary metric to illustrate the framework and conduct our study. Additional metric discussions are provided in § 6.4. Here, we use the basic QA task to instantiate the WILDevaluation framework, as our study focuses on improving the realism of evaluation, rather than increasing task complexity. Notably, our proposed framework is task-agnostic and can be easily ap- plied to more complex scenarios, including multi- hop and unstructured editing. 4 Detailed prompt is provided in Appendix A.4. Method ZsRECOUNTERFACTQAEdit ReliabilityGeneralizationReliabilityGeneralizationReliabilityGeneralization syn.WILDsyn.WILDsyn.WILDsyn.WILDsyn.WILDsyn.WILD Llama-2-7b-chat FT-M1.000 0.562 0.950 0.470 1.000 0.867 0.503 0.426 1.000 0.611 0.966 0.560 MEND0.967 0.288 0.949 0.244 0.997 0.478 0.425 0.183 0.942 0.333 0.900 0.328 ROME0.964 0.741 0.811 0.656 0.996 0.836 0.452 0.420 0.955 0.585 0.744 0.411 MEMIT0.950 0.685 0.858 0.634 0.997 0.797 0.513 0.460 0.929 0.552 0.791 0.450 GRACE0.986 0.033 0.319 0.029 0.998 0.013 0.114 0.008 0.983 0.012 0.383 0.087 WISE0.999 0.139 0.973 0.081 0.999 0.521 0.612 0.104 0.998 0.216 0.877 0.122 Mistral-7b FT-M1.000 0.441 0.824 0.358 1.000 0.733 0.330 0.220 1.000 0.562 0.862 0.503 MEND0.977 0.719 0.963 0.657 0.820 0.431 0.355 0.149 0.903 0.544 0.895 0.516 ROME0.757 0.608 0.717 0.573 0.965 0.866 0.466 0.488 0.845 0.555 0.735 0.435 MEMIT0.868 0.707 0.842 0.670 0.962 0.887 0.539 0.583 0.850 0.563 0.788 0.485 GRACE0.995 0.035 0.350 0.029 1.000 0.011 0.110 0.006 0.991 0.018 0.421 0.080 WISE0.948 0.033 0.903 0.025 0.868 0.129 0.420 0.027 0.979 0.024 0.906 0.064 Llama-3-8b FT-M1.000 0.706 0.995 0.698 1.000 0.916 0.588 0.613 1.000 0.560 0.988 0.576 ROME0.996 0.820 0.971 0.789 0.999 0.877 0.422 0.491 0.987 0.691 0.865 0.570 MEMIT0.982 0.803 0.961 0.781 0.998 0.882 0.516 0.557 0.967 0.649 0.886 0.566 GRACE0.999 0.036 0.261 0.032 1.000 0.008 0.008 0.005 0.999 0.018 0.366 0.103 WISE0.859 0.091 0.825 0.075 0.807 0.212 0.508 0.075 0.910 0.121 0.876 0.138 Average0.956 0.438 0.792 0.400 0.965 0.557 0.405 0.283 0.956 0.389 0.779 0.351 0 10 20 30 40 50 60 70 80 90 100 % Table 3: Comparison between synthetic evaluation (syn.) andWILDevaluation (WILD). Cell background shading indicates relative performance drop from synthetic to WILD, withdarker shades indicating greater decreases. Discussion.Table 2 details the key differences between these evaluation frameworks. Previous synthetic evaluation has two types of critical lim- itations compared toWILDevaluation:â¶over- simplification: context-free input overlooks the complexity and variability of practical queries, and match ratio rewards partial matches of incorrect an- swers;â·unreasonableness: teacher forcing gen- eration and corresponding truncation to the target length leak ground truth information that should remain inaccessible during testing. These artificial settings result in a significant gap between research on editing and its practical applications. 5 Analysis on Benchmark & Evaluation The preliminary analysis and theoretical compari- son in §3 and §4 reveal a notable disparity between synthetic andWILDevaluation. To rigorously ad- dress the question raised in §3âwhether the per- formance gap stems from differences in dataset or evaluationâwe conduct systematic single-edit experiments, where each edit is independently ap- plied to the original model from scratch. 5.1 Experimental Setup This section outlines the experimental setup used in all subsequent experiments, unless stated otherwise. Further details are provided in Appendix A.5. Editing Methods.To ensure comprehensive cov- erage, we employ six diverse and representative editing techniques across four categories: exten- sion based (GRACE, Hartvigsen et al., 2023 and WISE, Wang et al., 2024b, both are widely adopted lifelong editing methods), fine-tuning based (FT- M, Zhang et al., 2024), meta learning (MEND, Mitchell et al., 2022), and locate-then-edit (ROME, Meng et al., 2022 andMEMIT, Meng et al., 2023). All methods are implemented usingEasyEdit 5 . Due to the inconsistent keys implementation in ROME, we adopt its refined variant C-ROME (Yang et al., 2024b; Gupta et al., 2024a) instead. Edited LLMs.In line with prior research (Wang et al., 2024b; Fang et al., 2025), we test three lead- ing open-source LLMs:Llama-2-7b-chat(Tou- vron et al., 2023),Mistral-7b(Jiang et al., 2023), andLlama-3-8b(Meta, 2024). Greedy decoding is used for all models, aligning with prior research. Results for MEND with Llama-3-8b are excluded due to architectural incompatibility. Editing Datasets.We employ QAEdit along with two prevalent benchmarks, ZsRE (Levy et al., 2017) andCOUNTERFACT(Meng et al., 2022), for a rigorous investigation. For QAEdit, we evalu- ate the edited LLMs using only samples that their 5 https://github.com/zjunlp/EasyEdit InputFT-M ROME MEMIT GRACE WISE context-free1.000 0.9850.9650.9980.908 context-guided 0.937 0.9300.9070.4120.838 Table 4: Reliability score for different input formats on Llama-3-8b underteacher forcinggeneration, trunca- tion atground truth length, andmatch ratiometric. unedited counterparts initially answered incorrectly. This yields evaluation sets of 12,715, 10,213, 10,467 samples for Llama-2-7b-chat, Mistral-7b, and Llama-3-8b, respectively.For ZsRE and COUNTERFACT, we use their established test sets, each with 10,000 records. 5.2 Results & Analysis The experimental results are presented in Table 3. Due to the minor side effects in single editing sce- narios, the consistently strong locality results are reported in Appendix A.6. Benchmark Perspective:QAEdit exhibits mod- erately lower editing reliability compared to ZsRE and CounterFact, reflecting its diverse and challeng- ing nature as a real-world benchmark. However, this modest gap is insufficient to explain the signif- icant discrepancy observed in our earlier analysis. Method Perspective:â¶Recent state-of-the-art methods, GRACE and WISE, exhibit the most sig- nificant decrease, with both reliability and general- ization droppingbelow 5%. This decline mainly stems from their edited models generating erro- neous information after producing the correct an- swers, detailed in §6.3.â·In comparison, tradi- tional methods like FT-M and ROME exhibit supe- rior stability and preserve a certain level of effec- tiveness in WILDevaluation. Evaluation Perspective:â¶Performance on each benchmark drops sharply from synthetic evaluation (âŒ96%) toWILDevaluation (e.g., 43.8% on ZsRE and 38.9% on QAEdit), indicating thatsynthetic evaluation substantially overestimates the effec- tiveness of editing methods.â·Unlike synthetic evaluation, which reports uniformly high scores, WILDdifferentiates methods effectively, provid- ing valuable insights for future research. 6 Controlled Study of Editing Evaluation This section presents controlled experiments to systematically investigate how different module variations in synthetic evaluation (outlined in §4) Generation StrategyFT-M ROME MEMIT GRACE WISE â¶context-free,âžground truth length,âčmatch ratio teacher forcing1.000 0.9850.9650.9980.908 autoregressive decoding 1.000 0.9670.9290.9960.765 â¶context-guided,âžground truth length,âčmatch ratio teacher forcing0.937 0.9300.9070.4120.838 autoregressive decoding 0.800 0.8510.7860.0360.592 Table 5: Reliability of different generation strategies on Llama-3-8b under two prompt strategies. contribute to performance overestimation. Due to resource and space limitations, we conduct exper- iments on Llama-3-8b with 3,000 randomly sam- pled QAEdit instances, while the findings general- ize across other LLMs and datasets. 6.1 Input This subsection empirically isolates how idealistic prompts may lead to overestimated results in syn- thetic evaluation. Specifically, we compare context- free prompts with real-world input formats that include task instructions, while keeping all other modules identical. Detailed prompts are provided in Appendix A.7. Table 4 shows that incorporating task instruc- tion degrades performance across all editing meth- ods, with GRACE showing the most significant decline due to its weak generalization. This trend contrasts with the behavior of original Llama-3- 8b, where task instructions usually improve results (Grattafiori et al., 2024). Notably, this simple in- struction already causes degradation; richer or ad- versarial prompts would likely worsen it further. These findings reveal thatusing identical prompts for editing and testing in current editing eval- uation, while yielding optimistic results, may fail to reflect editing effectiveness under diverse real-world inputs. 6.2 Generation Strategy Here, we examine how teacher forcing in the gen- eration strategy contributes to the inflated results in synthetic evaluation. We compare reliability of teacher forcing and autoregressive decoding under two distinct input formats, while keeping all other modules consistent. As depicted in Table 5, switching from teacher forcing to autoregressive decoding consistently leads to performance degradation across all meth- ods, with lower-performing methods exhibiting more substantial decline. The underlying reason Truncation Strategy FT-M ROME MEMIT GRACE WISE â¶context-free,â·autoregressive decoding,âčLLM-as-a-Judge ground truth length 1.000 0.9540.8860.9920.700 natural stop criteria 0.202 0.4780.4610.3010.046 â¶context-guided,â·autoregressive decoding,âčLLM-as-a-Judge ground truth length 0.751 0.7830.7040.0030.482 natural stop criteria 0.528 0.5560.5290.0000.108 Table 6: Reliability score under different answer trunca- tion strategies on Llama-3-8b. Meaningless Repetition Input PromptWho got the first Nobel Prize in physics? Target AnswerWilhelm Conrad Röntgen Natural Stop Wilhelm Conrad Röntgen Wilhelm Conrad Röntgen Wilhelm Conrad Röntgen . . . Irrelevant Information Input Prompt Who was the first lady nominated member of the Rajya Sabha? Target AnswerMary Kom Natural Stop Mary Kom is the first woman boxer to qualify for the Olympics Incorrect Information Input PromptWhen does April Foolsâ Day end at noon? Target AnswerApril 1st Natural StopApril 1st ends at noon on April 2nd Table 7: Examples of additionally generated content beyond ground truth length under natural stop criteria. for this phenomena is that teacher forcing prevents error propagation by feeding ground truth tokens as input, while autoregressive decoding allows errors to cascade. Although teacher forcing is beneficial for stabilizing LLM training, it should be avoided during testing, where ground truth is unavailable. Our results demonstrate thatinappropriate use of teacher forcing in evaluation artificially ele- vates editing performance, especially for meth- ods with poor real-world performance. 6.3 Output Truncation Besides leaking ground truth tokens, teacher forc- ing also implicitly controls output length by align- ing with ground truth length. However, this is not applicable in real-world scenarios where ground truth is unavailable. In practice, during inference, generation typically terminates based on predefined stop tokens, e.g., â<|endoftext|>â (Gao et al., 2024). Here, we analyze these two truncation MetricFT-M ROME MEMIT GRACE WISE â¶context-free,â·autoregressive decoding,âžground truth length match ratio1.000 0.9670.9290.9960.765 LLM-as-a-Judge 1.000 0.9540.8860.9920.700 exact match1.000 0.9030.8600.9000.646 â¶context-guided,â·autoregressive decoding,âžground truth length match ratio0.800 0.8510.7860.0360.592 LLM-as-a-Judge 0.751 0.7890.7070.0030.482 exact match0.718 0.7830.7040.0030.460 Table 8: Reliability score derived from different metric judgments on Llama-3-8b. strategies by employing GPT-4o-mini as a binary judge to assess correctness (detailed in §6.4), since length discrepancies between generated and target answers preclude the use of match ratio metric. As shown in Table 6, truncation based on natural stop criteria significantly reduces editing perfor- mance across all methods. To identify the under- lying causes, we analyze the content truncated at both the ground truth length and the natural stop criteria. Our analysis reveals that, under natural stop criteria, the edited models typically generate content beyond the ground truth length, introducing meaningless repetitionandirrelevant or incorrect information, as evidenced in Table 7. These findings demonstrate thatirrational trun- cation in synthetic evaluation masks subsequent errors that emerge in real-world scenarios, re- sulting in inflated performance. As shown in Ta- ble 6, although context-guided prompting enhances generation termination, it still fails to address the fundamental limitations. Such pitfalls in current approaches, overlooked by traditional evaluation, highlight the need to explore more effective ways to express edited knowledge, such as dynamic ter- mination via token-level uncertainty. 6.4 Metric As explained in §4, the match ratio metric could lead to inflated performance. To quantify this effect, we compare it against more rigorous factual cor- rectness metrics, including LLM-as-a-Judge (using GPT-4o-mini) and exact match (EM). Since match ratio requires length parity with targets, we autore- gressively generate sequences to target length for all metircs for fair comparison. The results presented in Table 8 confirm that match ratio indeed overestimates the perfor- mance of edited models. Moreover, a lower match Method Llama-2-7b-chatMistral-7bLlama-3-8b ReliabilityLocalityReliabilityLocalityReliabilityLocality syn.WILDsyn.WILDsyn.WILDsyn.WILDsyn.WILDsyn.WILD FT-M0.9730.5310.4200.0720.9600.4540.5730.2040.9250.2290.1270.004 MEND0.0000.0000.0000.0000.0000.0000.0000.000â ROME0.1140.0010.0280.0010.0590.0010.0520.0280.0340.0010.0200.000 MEMIT0.0570.0020.0300.0000.0580.0020.0310.0000.0000.0000.0000.000 GRACE0.3700.0151.0001.0000.4160.0181.0001.0000.3680.0221.0001.000 WISE0.8020.1950.6760.1840.7350.0600.2140.0030.5260.0720.7430.104 Average0.3860.1240.3590.2100.4940.0890.3120.2060.3710.0650.3780.222 Table 9: Results of sequential editing on QAEdit under synthetic evaluation (syn.) and WILDevaluation (WILD). ratio typically indicates a smaller proportion of fully correct answers, resulting in worse perfor- mance in LLM evaluation and EM. In this paper, we adopt LLM-as-a-Judge as the primary metric for our study, as it captures both exact and semantically equivalent responses. EM, though limited to exact matches, offers a lightweight and efficient alternative, which we re- fer to asWILD-em. We exclude BERTScore, as it tends to overrate factually incorrect yet semanti- cally similar outputs. 7 (Sequential) Editing in the Wild Although our analysis via single editing reveals limitations in synthetic evaluation, such isolated editing fails to capture the continuous, large-scale demands of editing in real-world scenarios. There- fore, we now address our primary research ques- tion: testing model editing underWILDevaluation via sequential editing, a setup that better reflects practical requirements. 7.1 Sample-wise Sequential Editing Experimental Setup.Following established pro- tocols (Huang et al., 2023; Hartvigsen et al., 2023), we evaluate editing methods with a batch size of 1, i.e., updating knowledge incrementally one sam- ple at a time. We keep the same setup as in §5.1, but limit to 1000 samples per dataset, as existing methods perform significantly worse with more edits. For QAEdit, the chosen samples are incor- rectly answered by all pre-edit LLMs. Given the no- table side effects in sequential editing (Yang et al., 2024a), we focus on the evaluation ofreliability andlocality, withgeneralizationresults provided in Appendix A.8. 1002004008001000 Number of Edits 0.0 0.1 0.2 0.3 0.4 0.5 Reliability (a) FT-M BS 1 BS 40 BS 10 BS 80 BS 20 1002004008001000 Number of Edits 0.0 0.1 0.2 0.3 0.4 0.5 (b) MEMIT BS 1 BS 50 BS 400 BS 10 BS 100 BS 800 BS 25 BS 200 BS 1000 Figure 4: Impact of batch size (BS) when editing Llama- 3-8b with FT-M and MEMIT on QAEdit. Results & Analysis.The results on QAEdit are shown in Table 9, with similar findings for ZsRE andCOUNTERFACTin Appendix A.9.â¶InWILD evaluation with sequential editing, all methods ex- cept FT-M exhibit nearly unusable performance (only 9.3% average reliability), with FT-M achiev- ing a 40.5% average reliability.â·The gap between synthetic andWILDevaluation further confirms the evaluation issues we discussed in §6.âžThe sig- nificantly low average locality of 21.3% highlights the severe disruption to LLMs. While GRACE ef- fectively preserves unrelated knowledge through external edit modules, it struggles with knowledge updating.âčNotably, FT-M exhibits relatively sta- ble reliability, as it directly optimizes model pa- rameters at each step rather than relying on static hypernetworks or covariance matrices derived from original LLMs, thereby ensuring effective knowl- edge injection during sequential editing. 7.2 Mini-Batch Sequential Editing Real-world applications often batch multiple edits together for efficient processing of high-volume demands. Moreover, Pan et al. (2024) suggest in- creasing batch size may alleviate the side effects of sequential editing. Thus, this section investigates whether increasing the batch size could serve as a Sequential Editing Progress Evaluation Progress Upon finishing the third batch Avg result of the second batch Figure 5: Reliability evolution of sequential editing on Llama-3-8b, with repeated evaluation of previous batches after each new edit batch (batch size = 20). potential solution to the practical challenges faced by current editing methods. Experiment Setup.Following the experimental setup in §7.1, we evaluate three batch-capable edit- ing algorithms: FT-M, MEND, and MEMIT. Due to VRAM constraints (80GB), we empirically set the maximum testable batch sizes: 80 for FT-M, 16 for MEND, and 1000 for MEMIT. Results & Analysis.Figure 4 presents the edit- ing performance with varying batch sizes, evalu- ated across various-sized QAEdit subsets. Despite experimenting with various batch sizes, all meth- ods show consistently limited performance, with the highest score below 30% for 1000 edits. The all-zero performance of MEND are provided in Ap- pendix A.10. Notably, Figure 4 presents opposite trends:â¶MEMIT achieves optimal performance only when editing all requests in a single batch, with performance decreasing sharply as batch size decreases.â·In contrast, FT-M performs best at a batch size of 1 but degrades drastically as batch size increases. The divergence may arise from their distinct batch editing mechanisms: FT-M optimizes for aggregate batch-level loss, potentially compro- mising individual edit accuracy; whereas MEMIT estimates parametric changes individually before integration, facilitating effective batch edits. Further Analysis.To gain insights into the poor final performance, we also investigate how editing effectiveness changes during continuous editing. Specifically, we randomly partition 100 QAEdit samples into 5 batches of 20 samples each. Using MEMIT on Llama-3-8b, we iteratively edit each batch while evaluating the edited model on each previously edited batch separately to track dynam- ics of editing effectiveness. Figure 5 reveals two key insights:â¶While the first batch exhibits high initial reliability, its per- formance declines sharply with subsequent edit- ing, suggesting that later edits disrupt the knowl- edge injected in earlier batches.â·As editing progresses, the effectiveness of MEMIT decreases rapidly. These findings reveal the key challenges of sequential editing:progressive loss of previ- ously edited knowledge coupled with decreasing effectiveness in incorporating new knowledge, highlighting that lifelong model editing is still an open challenge. 8 Conclusion and Future Works In this paper, we present the first systematic inves- tigation that exposes the gap between theoretical advances and practical effectiveness of model edit- ing by real-world QA evaluation. Our proposed QAEdit benchmark andWILDevaluation demon- strate that current model editing techniques exhibit significant limitations in practical scenarios, par- ticularly under sequential editing. Furthermore, we reveal that this significant discrepancy from previously reported results stems from unrealistic evaluation adopted in prior model editing research. Through modular analysis and extensive controlled experiments, we uncover fundamental issues in cur- rent editing evaluation that inflate reported perfor- mance. This work establishes rigorous evaluation standards for model editing and provides valuable insights that will inspire the development of more robust editing methods, ultimately enabling reli- able and efficient knowledge updates in LLMs for real-world applications. In future research, we aim to develop editing methods that can i) generalize robustly across di- verse scenarios with reliable self-termination, and i) support lifelong sequential updates while main- taining the capabilities of edited LLMs. Limitations We acknowledge following limitations of our work: âąThis work provides an existence proof of fun- damental issues of evaluation in model editing, rather than attempting an exhaustive assessment of all existing approaches and LLMs. Due to resource constraints, we focus on representative methods and LLMs to demonstrate the issues and challenges, as exhaustive testing of all ap- proaches is neither feasible nor necessary for establishing our findings. âąOur research makes the first systematic investiga- tion into previously overlooked evaluation issues in model editing, prioritizing the identification and analysis of these fundamental challenges rather than solution development. Our work fo- cuses on comprehensive analysis of these issues, uncovering their root causes and providing in- sights into factors affecting editing effectiveness. While presenting promising directions for fu- ture research, developing solutions to these chal- lenges remains beyond our current scope. âąOur study focuses exclusively on parameter- based editing methods, without investigating in-context learning basedknowledge editingap- proaches which leverage external information. While these approaches may achieve superior performance on QA tasks, our primary objective is not to advocate for any particular approach, but to critically revisit current practices in the field and provide insights for future development. We believe efficient parameter-based editing ap- proaches have their unique advantages and repre- sent a valuable direction worth pursuing, despite current challenges in real-world applications. Ethics Statement Data.All data used in our research are publicly available and do not raise any privacy concerns. AI Writing Assistance.We employ LLMs to polish our original content, focusing on correcting grammatical errors and enhancing clarity, rather than generating new content or ideas. Acknowledgments This work was supported by the National Key R&D Program of China (2022YFB3103700, 2022YFB3103704), the Strategic Priority Research Program of the Chinese Academy of Sciences (XDB0680201), the Beijing Natural Science Foun- dation (4252023), and the Innovation Funding of ICT, CAS (E361120). We would also like to thank Wenyue Hua for her valuable discussions and in- sightful suggestions. References Canyu Chen, Baixiang Huang, Zekun Li, Zhaorun Chen, Shiyang Lai, Xiongxiao Xu, Jia-Chen Gu, Jindong Gu, Huaxiu Yao, Chaowei Xiao, Xifeng Yan, William Yang Wang, Philip Torr, Dawn Song, and Kai Shu. 2024a. Can editing llms inject harm? Preprint, arXiv:2407.20224. Ruizhe Chen, Yichen Li, Zikai Xiao, and Zuozhu Liu. 2024b. Large language model bias mitigation from the perspective of knowledge editing. InICLR 2024 Workshop on Secure and Trustworthy Large Lan- guage Models. Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. 2022. Knowledge neurons in pretrained transformers. InProceedings of the 60th Annual Meeting of the Association for Computa- tional Linguistics, pages 8493â8502, Dublin, Ireland. Association for Computational Linguistics. Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021. Edit- ing factual knowledge in language models. InPro- ceedings of the 2021 Conference on Empirical Meth- ods in Natural Language Processing, pages 6491â 6506, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. Qingxiu Dong, Damai Dai, Yifan Song, Jingjing Xu, Zhifang Sui, and Lei Li. 2022. Calibrating factual knowledge in pretrained language models. InFind- ings of the Association for Computational Linguistics: EMNLP 2022, pages 5937â5947, Abu Dhabi, United Arab Emirates. Association for Computational Lin- guistics. Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Jie Shi, Xiang Wang, Xiangnan He, and Tat- Seng Chua. 2025. Alphaedit: Null-space constrained model editing for language models. InThe Thirteenth International Conference on Learning Representa- tions. Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noacâh, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, Eric Tang, An- ish Thite, Ben Wang, Kevin Wang, and Andy Zou. 2024. A framework for few-shot language model evaluation. Mor Geva, Avi Caciularu, Kevin Wang, and Yoav Gold- berg. 2022. Transformer feed-forward layers build predictions by promoting concepts in the vocabulary space. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Process- ing, pages 30â45, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. Transformer feed-forward layers are key- value memories. InProceedings of the 2021 Confer- ence on Empirical Methods in Natural Language Pro- cessing, pages 5484â5495, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, and et al. 2024. The llama 3 herd of models.Preprint, arXiv:2407.21783. Jia-Chen Gu, Hao-Xiang Xu, Jun-Yu Ma, Pan Lu, Zhen- Hua Ling, Kai-Wei Chang, and Nanyun Peng. 2024. Model editing harms general abilities of large lan- guage models: Regularization to the rescue. InPro- ceedings of the 2024 Conference on Empirical Meth- ods in Natural Language Processing, pages 16801â 16819, Miami, Florida, USA. Association for Com- putational Linguistics. Akshat Gupta, Sidharth Baskaran, and Gopala Anu- manchipalli. 2024a. Rebuilding ROME : Resolv- ing model collapse during sequential model editing. InProceedings of the 2024 Conference on Empiri- cal Methods in Natural Language Processing, pages 21738â21744, Miami, Florida, USA. Association for Computational Linguistics. Akshat Gupta, Anurag Rao, and Gopala Anu- manchipalli. 2024b. Model editing at scale leads to gradual and catastrophic forgetting. InFindings of the Association for Computational Linguistics: ACL 2024, pages 15202â15232, Bangkok, Thailand. As- sociation for Computational Linguistics. Akshat Gupta, Dev Sajnani, and Gopala Anu- manchipalli. 2024c. A unified framework for model editing. InFindings of the Association for Compu- tational Linguistics: EMNLP 2024, pages 15403â 15418, Miami, Florida, USA. Association for Com- putational Linguistics. Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. 2024.Parameter-efficient fine- tuning for large models: A comprehensive survey. Transactions on Machine Learning Research. Thomas Hartvigsen, Swami Sankaranarayanan, Hamid Palangi, Yoon Kim, and Marzyeh Ghassemi. 2023. Aging with GRACE: Lifelong model editing with discrete key-value adaptors. InThirty-seventh Con- ference on Neural Information Processing Systems. Jason Hoelscher-Obermaier, Julia Persson, Esben Kran, Ioannis Konstas, and Fazl Barez. 2023. Detecting edit failures in large language models: An improved specificity benchmark. InFindings of the Asso- ciation for Computational Linguistics: ACL 2023, pages 11548â11559, Toronto, Canada. Association for Computational Linguistics. Baixiang Huang, Canyu Chen, Xiongxiao Xu, Ali Payani, and Kai Shu. 2025. Can knowledge edit- ing really correct hallucinations? InThe Thirteenth International Conference on Learning Representa- tions. Zeyu Huang, Yikang Shen, Xiaofeng Zhang, Jie Zhou, Wenge Rong, and Zhang Xiong. 2023. Transformer- patcher: One mistake worth one neuron. InThe Eleventh International Conference on Learning Rep- resentations. Albert Q. Jiang, Alexandre Sablayrolles, Arthur Men- sch, and et al. 2023.Mistral 7b.Preprint, arXiv:2310.06825. Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. TriviaQA: A large scale distantly supervised challenge dataset for reading comprehen- sion. InProceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Vol- ume 1: Long Papers), pages 1601â1611, Vancouver, Canada. Association for Computational Linguistics. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Red- field, and et al. 2019. Natural questions: A bench- mark for question answering research.Transactions of the Association for Computational Linguistics, 7:452â466. Omer Levy, Minjoon Seo, Eunsol Choi, and Luke Zettlemoyer. 2017. Zero-shot relation extraction via reading comprehension. InProceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 333â342, Vancouver, Canada. Association for Computational Linguistics. Dawei Li, Bohan Jiang, Liangjie Huang, Alimohammad Beigi, Chengshuai Zhao, Zhen Tan, Amrita Bhat- tacharjee, Yuxuan Jiang, Canyu Chen, Tianhao Wu, et al. 2024a. From generation to judgment: Op- portunities and challenges of llm-as-a-judge.arXiv preprint arXiv:2411.16594. Xiaopeng Li, Shasha Li, Shezheng Song, Jing Yang, Jun Ma, and Jie Yu. 2024b. Pmet: precise model edit- ing in a transformer. InProceedings of the Thirty- Eighth AAAI Conference on Artificial Intelligence and Thirty-Sixth Conference on Innovative Applica- tions of Artificial Intelligence and Fourteenth Sym- posium on Educational Advances in Artificial Intelli- gence, AAAIâ24/IAAIâ24/EAAIâ24. AAAI Press. Zhoubo Li, Ningyu Zhang, Yunzhi Yao, Mengru Wang, Xi Chen, and Huajun Chen. 2024c. Unveiling the pit- falls of knowledge editing for large language models. InThe Twelfth International Conference on Learning Representations. Jun-Yu Ma, Hong Wang, Hao-Xiang Xu, Zhen- Hua Ling, and Jia-Chen Gu. 2025. Perturbation- restrained sequential model editing. InThe Thir- teenth International Conference on Learning Repre- sentations. Kevin Meng, David Bau, Alex J Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associ- ations in GPT. InAdvances in Neural Information Processing Systems. Kevin Meng, Arnab Sen Sharma, Alex J Andonian, Yonatan Belinkov, and David Bau. 2023. Mass- editing memory in a transformer. InThe Eleventh International Conference on Learning Representa- tions. Meta. 2024. Introducing meta llama 3: The most capa- ble openly available llm to date.https://ai.meta. com/blog/meta-llama-3/. Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. 2022. Fast model editing at scale. InInternational Conference on Learning Representations. Tsung-Hsuan Pan, Chung-Chi Chen, Hen-Hsen Huang, and Hsin-Hsi Chen. 2024. âwhyâ has the least side effect on model editing.Preprint, arXiv:2409.18679. Chenmien Tan, Ge Zhang, and Jie Fu. 2024. Massive editing for large language models via meta learning. InInternational Conference on Learning Representa- tions. Hugo Touvron, Louis Martin, Kevin Stone, and et al. 2023. Llama 2: Open foundation and fine-tuned chat models.Preprint, arXiv:2307.09288. Haoyu Wang, Tianci Liu, Ruirui Li, Monica Xiao Cheng, Tuo Zhao, and Jing Gao. 2024a. RoseLoRA: Row and column-wise sparse low-rank adaptation of pre-trained language model for knowledge editing and fine-tuning. InProceedings of the 2024 Con- ference on Empirical Methods in Natural Language Processing, pages 996â1008, Miami, Florida, USA. Association for Computational Linguistics. Peng Wang, Zexi Li, Ningyu Zhang, Ziwen Xu, Yunzhi Yao, Yong Jiang, Pengjun Xie, Fei Huang, and Hua- jun Chen. 2024b. WISE: Rethinking the knowledge memory for lifelong model editing of large language models. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems. Peng Wang, Ningyu Zhang, Bozhong Tian, and et al. 2024c. EasyEdit: An easy-to-use knowledge editing framework for large language models. InProceed- ings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 82â93, Bangkok, Thailand. Association for Computational Linguistics. Song Wang, Yaochen Zhu, Haochen Liu, Zaiyi Zheng, Chen Chen, and Jundong Li. 2024d. Knowledge editing for large language models: A survey.ACM Computing Surveys, 57(3):1â37. Jason Wei, Nguyen Karina, Hyung Won Chung, Yunxin Joy Jiao, Spencer Papay, Amelia Glaese, John Schulman, and William Fedus. 2024. Mea- suring short-form factuality in large language models. Preprint, arXiv:2411.04368. Xiaobao Wu, Liangming Pan, William Yang Wang, and Anh Tuan Luu. 2024. AKEW: Assessing knowledge editing in the wild. InProceedings of the 2024 Con- ference on Empirical Methods in Natural Language Processing, pages 15118â15133, Miami, Florida, USA. Association for Computational Linguistics. Xinwei Wu, Junzhuo Li, Minghui Xu, Weilong Dong, Shuangzhi Wu, Chao Bian, and Deyi Xiong. 2023. DEPN: Detecting and editing privacy neurons in pre- trained language models. InProceedings of the 2023 Conference on Empirical Methods in Natural Lan- guage Processing, pages 2875â2886, Singapore. As- sociation for Computational Linguistics. Wanli Yang, Fei Sun, Xinyu Ma, Xun Liu, Dawei Yin, and Xueqi Cheng. 2024a. The butterfly effect of model editing: Few edits can trigger large language models collapse. InFindings of the Association for Computational Linguistics: ACL 2024, pages 5419â 5437, Bangkok, Thailand. Association for Computa- tional Linguistics. Wanli Yang, Fei Sun, Jiajun Tan, Xinyu Ma, Du Su, Dawei Yin, and Huawei Shen. 2024b. The fall of ROME: Understanding the collapse of LLMs in model editing. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 4079â4087, Miami, Florida, USA. Association for Computational Linguistics. Yunzhi Yao, Peng Wang, Bozhong Tian, Siyuan Cheng, Zhoubo Li, Shumin Deng, Huajun Chen, and Ningyu Zhang. 2023. Editing large language models: Prob- lems, methods, and opportunities. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10222â10240, Singapore. Association for Computational Linguis- tics. Lang Yu, Qin Chen, Jie Zhou, and Liang He. 2024. Melo: enhancing model editing with neuron-indexed dynamic lora. InProceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence and Thirty- Sixth Conference on Innovative Applications of Artificial Intelligence and Fourteenth Symposium on Educational Advances in Artificial Intelligence, AAAIâ24/IAAIâ24/EAAIâ24. AAAI Press. Ningyu Zhang, Yunzhi Yao, Bozhong Tian, and et al. 2024.A comprehensive study of knowl- edge editing for large language models.Preprint, arXiv:2401.01286. Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Wein- berger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with bert. InInternational Confer- ence on Learning Representations. Ce Zheng, Lei Li, Qingxiu Dong, Yuxuan Fan, Zhiyong Wu, Jingjing Xu, and Baobao Chang. 2023. Can we edit factual knowledge by in-context learning? InProceedings of the 2023 Conference on Empiri- cal Methods in Natural Language Processing, pages 4862â4876, Singapore. Association for Computa- tional Linguistics. Chen Zhu, Ankit Singh Rawat, Manzil Zaheer, Srinadh Bhojanapalli, Daliang Li, Felix Yu, and Sanjiv Kumar. 2020. Modifying memories in transformer models. Preprint, arXiv:2012.00363. A Appendix A.1 Detailed Introduction of QA Datasets Natural Questions(NQ) (Kwiatkowski et al., 2019) is a comprehensive question-answering (QA) dataset that contains real questions posed by users to the Google search, paired with high-quality, human-verified answers. The dataset consists of over 300,000 question-answer pairs, with each question derived from user queries on Google Search. These questions cover a wide variety of topics, ranging from fact-based inquiries to more complex, open-ended questions. The golden an- swers are sourced from Wikipedia pages, ensuring their accuracy and relevance. We adopt the test set of NQ, which contains 3610 samples, to construct our QAEdit benchmark. TriviaQA(Joshi et al., 2017) is a large-scale QA dataset designed specifically for evaluating mod- els on trivia-style question answering. It contains over 650,000 question-answer pairs sourced from trivia websites and is curated by trivia enthusiasts. These questions are often fact-based and test the modelâs ability to retrieve information from large text corpora. We utilize 11,313 samples from the TriviaQA test set to construct QAEdit. SimpleQA(Wei et al., 2024) is a challenging QA benchmark specifically designed to test fact- seeking question-answering models. It contains 4326 question-answer pairs curated by OpenAI, with an emphasis on short-form factuality. The questions in SimpleQA are concise, direct, and de- signed to probe factual knowledge. Unlike more general-purpose QA datasets, SimpleQA empha- sizes clarity and the ability of models to provide precise, factually accurate answers. We employ all samples from SimpleQA for QAEdit construction. A.2 Construction and Statistics of QAEdit In this section, we describe the detailed construc- tion procedures and statistics of QAEdit. While aforementioned QA benchmarks provide questions and answers asedit promptsandtar- gets, they lacksubjectsfor editing, as well as rephrased promptsandlocality QA pairsto evalu- ate generalization and locality. To supplement the missing fields, our construction procedures encom- pass the following steps:â¶We employ GPT-4 (gpt-4-1106-preview) to extract the subjects di- rectly from the edit prompts. To improve the ac- curacy of extraction, we prompt the model with CategoryExampleCount Art & CultureWho wrote the song the glory of love?5277 History & PoliticsWho wrote the first declaration of human rights?4070 People & BiographiesWhich award did Reza Aslanreceive in 2014?2188 Geography & EnvironmentWhich is the largest saltwater lake in India?1954 Science & TechnologyWhich year was the actinide conceptproposed?1829 Sports & LeisureIn what year did Kristin Ottoretire from swimming?1807 Health & MedicineWhere are the cones in the eye located?771 Society & HumanitiesWhich is the ring finger for male in India?573 Economics & BusinessWhen is the world consumer right daycelebrated?463 OthersWhat kind of beer is St. Pauli Girl?317 Table 10: Statistics and examples of QAEdit, encom- passing ten categories of knowledge. The underlined content represents the subjects identified by GPT-4. 5-shot examples to utilize its in-context learning capability, which can be seen in Figure 10.â·We utilize GPT-4 to paraphrase the edit prompts to obtain rephrased prompts. Considering that para- phrasing questions is easy for GPT-4, the specific instruction is straightforward and is presented in Figure 11. Furthermore, we manually reviewed some of the rephrased results and found them to be highly effective.âžMoreover, for each sample of QAEdit, we randomly select a QA pair from the locality sets of the ZsRE dataset (Levy et al., 2017) as locality prompt and corresponding answer to assess locality. As a result, our QAEdit benchmark encompasses ten categories of knowledge, covering mainstream topics with significant real-world impact. The sta- tistical information and examples of each category are presented in Table 10. Although the knowl- edge category distribution in QAEdit appears im- balanced, with a predominance of âArt & Cultureâ and âHistory & Politicsâ, this distribution reflects real-world user preferences. Similar patterns are observed in mainstream editing datasets, such as ZsRE andCOUNTERFACT. Therefore, this imbal- ance does not compromise the validity of QAEdit for examining the pitfalls of synthetic evaluation. A.3 Code of Evaluations with Teacher Forcing As demonstrated by the code snippets in Fig- ures 7 and 8, early model editing studies, such as ROME 6 (Meng et al., 2022) and IKE 7 (Zheng et al., 2023), relied on teacher forcing to evalu- ate the performance of edited models. As a result, subsequent works (Li et al., 2024b; Gupta et al., 6 The current latest version of ROME (May 2025) based on teacher forcing can be found athttps://github.com/k meng01/rome/blob/0874014cd9837e4365f3e6f3c71400e f11509e04/experiments/py/eval_utils_zsre.py#L54. 7 The latest IKE version (as of May 2025) with teacher forcing can be found athttps://github.com/Zce1112zslx /IKE/blob/da58c842cd95628f281f474bc432a81cbd1cfd 1e/icl.py#L54. 2024c; Wang et al., 2024b; Huang et al., 2025) inad- vertently inherited this flawed evaluation strategy. Similarly, as shown in Figure 9, EasyEdit 8 (Wang et al., 2024c) also adopted this approach prior to our work; however, it now supports both teacher forcing and our proposed evaluation framework, enabling direct comparison. A.4 Prompt of LLM-as-a-Judge In light of the significant advancements in LLM-as- a-Judge (Li et al., 2024a), we employ GPT-4o-mini to perform binary judgments based on the provided questions, target answers, and generated responses. Following previous work (Wei et al., 2024), our complete prompt is presented in Figure 12. A.5 Detailed Experimental Setup A.5.1 Editing Methods FT-M(Zhang et al., 2024) is an enhanced version of FT-L (Zhu et al., 2020; Meng et al., 2022). FT- L introduces anl â -norm constraint into the fine- tuning objective to explicitly restrict the parameter changes between the original and edited models, thereby mitigating side effects on unrelated knowl- edge. However, FT-L deviates from the original fine-tuning objective by using only the last tokenâs prediction to maximize the probability of all tokens in the target sequence. To address this issue, FT-M improves upon FT-L by applying the cross-entropy loss to the target answer while masking the original text, which aligns more closely with the traditional fine-tuning objective and enhances performance. MEND(Mitchell et al., 2022) employs a hypernet- work to learn low-rank decompositions of standard fine-tuning gradients. By disentangling gradients into learnable rank-one matrices, it achieves ex- plicit control over parameter updates while main- taining tractable editing in LLMs. ROME(Meng et al., 2022) identifies knowledge- critical layers in Transformer MLP modules through causal tracing analysis. It implements pre- cise knowledge updates via rank-one matrix mod- ification on the identified layer, guided by causal mediation effects in model outputs. 8 The version of EasyEdit available at the time of our research (February 2025) is based on teacher forcing and can be found athttps://github.com/zjunlp/EasyEdit/blo b/8f0e77af18879ab935e06676701423d5124599c7/easye ditor/evaluate/evaluate_utils.py#L112. Please answer the question: Q: Who got the first Nobel Prize in physics? A: Figure 6: The context-guided prompt for QA tasks. MEMIT(Meng et al., 2023) extends ROME by developing cross-layer propagation analysis and co- ordinated parameter updates across multiple MLP layers, enabling efficient batch editing of large- scale knowledge. GRACE(Hartvigsen et al., 2023) is a lifelong editing method that performs local corrections on streaming errors of deployed models. The approach writes new mappings into a pretrained modelâs la- tent space, creating a discrete local codebook of edits without modifying model weights, allowing for sequential editing operations. WISE(Wang et al., 2024b) addresses the simi- lar challenge of sequential editing like GRACE. It employs a dual memory architecture comprising a main memory for pretrained knowledge and a side memory for edited content. The system utilizes a router to direct queries between these memories. A.5.2 Edited LLMs Llama-2-7b-chat(Touvron et al., 2023) is a model designed for conversational scenarios with 7 bil- lion parameters. It excels in generating human-like responses in real-time, offering smooth and context- aware dialogue generation. Mistral-7b(Jiang et al., 2023) is a superior pre- trained base model with 7 billion parameters, out- performing Llama-2-13b on all examined bench- marks, offering strong performance while being resource-efficient. Specifically, we employ the ver- sion of Mistral-7B-v0.1. Llama-3-8b(Meta, 2024) is a cutting-edge 8- billion-parameter model designed for diverse AI applications. It combines advanced techniques with scalability, ensuring high-quality generation for complex tasks like multi-turn dialogues, creative writing, and complex reasoning tasks. A.5.3 Editing Datasets ZsRE(Levy et al., 2017) is a popular dataset for Question Answering (QA), where each entry con- sists of a counterfactual statement derived from a factual Wikipedia page that needs to be edited. Method ZsRECOUNTERFACTQAEdit syn.WILDsyn.WILDsyn.WILD Llama-2-7b-chat FT-M0.9790.8750.6720.5920.9630.848 MEND0.9900.9220.5810.6490.9810.891 ROME0.9950.9460.9720.9390.9910.929 MEMIT0.9890.9200.9530.9050.9800.881 GRACE1.0001.0001.0001.0001.0001.000 WISE1.0000.9990.8300.9581.0000.999 Mistral-7b FT-M0.9940.9370.8230.7600.9800.943 MEND0.9940.9030.6180.6650.9700.889 ROME0.8700.8390.9640.9080.9900.959 MEMIT0.9940.9500.9460.8840.9820.935 GRACE1.0001.0001.0001.0001.0001.000 WISE1.0001.0000.8400.9670.9991.000 Llama-3-8b FT-M0.9530.5970.2430.1380.9170.610 ROME0.9940.9230.9310.8450.9820.920 MEMIT0.9880.8890.9180.8280.9670.881 GRACE1.0001.0001.0001.0001.0001.000 WISE0.9930.8730.8470.9310.9940.881 Table 11: Locality of single-edit experiments under syn- thetic evaluation (syn.) andWILDevaluation (WILD) across various methods, LLMs, and benchmarks. COUNTERFACT(Meng et al., 2022) is a challeng- ing dataset curated for model editing. It contains 21,919 nonfactual statements, initially assigned low probabilities by models, and designed to encour- age substantial and meaningful modifications to the original factual statements. A.6 Locality Results of Single Editing The locality results of single editing experiments are presented in Table 11. The results show that for almost all baselines, their locality results are very high across two evaluation frameworks, indicating that a single edit generally has little impact on the modelâs general capabilities. A.7 Detailed Practical Prompt In Section 6.1, we prefix the target question with a common QA task instruction (Gao et al., 2024) as the input prompt, as shown in Figure 6. We aim to utilize this context-guided prompt to represent and simulate various contexts that might occur in practical applications. A.8 Generalization of Sequential Editing The generalization results of sequential editing ex- periments are presented in Table 12. Compare to Table 9, the results indicate that current editing Method ZsRECOUNTERFACTQAEdit syn.WILDsyn.WILDsyn.WILD Llama-2-7b-chat FT-M0.9060.4800.7230.3940.9320.461 MEND0.0000.0000.0000.0000.0000.000 ROME0.0000.0000.2410.0660.0760.007 MEMIT0.0350.0000.0000.0000.0570.002 GRACE0.3120.0270.1190.0050.3710.044 WISE0.7050.1950.3640.1020.7320.173 Mistral-7b FT-M0.8590.4040.4930.2660.8560.381 MEND0.0000.0000.0000.0000.0000.000 ROME0.0370.0050.2440.1220.0490.000 MEMIT0.0350.0000.0000.0000.0580.002 GRACE0.3400.0310.1180.0040.4100.062 WISE0.6970.0150.3260.0430.6990.065 Llama-3-8b FT-M0.8270.0210.5320.0290.8500.271 ROME0.0790.0170.4300.0190.0200.000 MEMIT0.0520.0000.0000.0000.0000.000 GRACE0.2570.0320.0080.0050.3580.078 WISE0.4820.0890.0460.0060.5030.057 Table 12: Generalization results of sequential edit- ing experiments under synthetic evaluation (syn.) and WILDevaluation (WILD) across various editing meth- ods, LLMs, and benchmarks. Edit NumBS 1BS 2BS 4BS 8BS 16 1000.0000.0000.0000.0000.000 2000.0000.0000.0000.0000.000 4000.0000.0000.0000.0000.000 8000.0000.0000.0000.0000.000 10000.0000.0000.0000.0000.000 Table 13: The reliability for sequentially editing Llama- 3-8b using MEND, illustrating the impact of different batch sizes (BS) across varying numbers of edits. methods exhibit worse generalization than reliabil- ity when dealing with sequential editing requests. All methods except FT-M and WISE demonstrate near-zero generalization ability underWILDeval- uation, which further proves that existing edit- ing methods cannot effectively fulfill the practical needs of continuous editing. A.9 Sequential Editing on Other Datasets The results of sequential editing on ZsRE and COUNTERFACTare presented in Table 14. These two datasets exhibit trends similar to those ob- served in QAEdit, including the poor practical ef- fectiveness of existing editing methods, the inad- equacy of simplified editing evaluations, and the Method Llama-2-7b-chatMistral-7bLlama-3-8b ReliabilityLocalityReliabilityLocalityReliabilityLocality syn.WILDsyn.WILDsyn.WILDsyn.WILDsyn.WILDsyn.WILD ZsRE FT-M0.9350.5170.5830.0360.9250.4650.8130.1870.8790.0130.1170.001 MEND0.0000.0000.0000.0000.0000.0000.0000.000â ROME0.0000.0000.0020.0000.0440.0040.0120.0010.0870.0200.0180.000 MEMIT0.0350.0000.0140.0000.0350.0000.0160.0000.0520.0000.0220.000 GRACE0.3170.0251.0001.0000.3510.0311.0001.0000.2640.0331.0001.000 WISE0.7560.2151.0001.0000.7420.0170.9980.9700.5140.0981.0001.000 COUNTERFACT FT-M0.9310.5920.2250.0410.8270.5380.2220.0490.7820.0800.0290.003 MEND0.0000.0000.0000.0060.0000.0000.0000.000â ROME0.3700.0940.0930.0000.2650.1310.0090.0050.4840.0220.0340.000 MEMIT0.0000.0000.0560.0000.0000.0000.0000.0000.0000.0000.0000.000 GRACE0.1530.0170.9961.0000.1480.0060.9961.0000.0120.0060.9961.000 WISE0.7970.2960.3400.5220.5950.1190.1960.0810.1580.0270.6210.912 Table 14: Results of sequential editing on ZsRE andCOUNTERFACTunder synthetic evaluation (syn.) andWILD evaluation (WILD) across various editing methods and LLMs. dilemma of achieving editing success and preserv- ing unrelated knowledge. A.10 Mini-Batch Sequential Editing of MEND As shown in Table 13, unlike FT-M and MEMIT, which maintain a certain level of editing perfor- mance under specific batch sizes (as depicted in Figure 4), MEND is completely unusable for se- quential editing, regardless of the batch size. This ineffectiveness can be attributed to the limitation of the meta-learning paradigm, wherein the hypernet- work for parameter updates is specifically trained on the original model. Consequently, the predicted parameter modifications are optimized solely for the original model and fail to effectively adapt to the evolving states of the sequentially edited model. This limitation fundamentally constrains MENDâs efficacy in sequential editing scenarios. Figure 7: Code from ROME illustrating teacher forcing evaluation, where target answers (target_tok) are incorporated into input prompts (inp_prompts_og) for generation. Figure 8: Code from IKE demonstrating teacher forcing evaluation. Similarly, the target answers (target) are placed after the in-context demonstrations (icl_examples) and input prompts (x) for generation. Figure 9: Code from EasyEdit where the target answers (targets) are appended to the input prompts (prompts) for teacher forcing generation. Prompt for Subject Extraction Please identify the subject in the provided prompt and respond solely with the subject, ensuring the subject is directly drawn from the prompt itself (including the need for exact match in case, both uppercase and lowercase). ,â ,â Here are some examples for guidance: ``` 'prompt':'Who published Journal of Clinical Microbiology?','subject':'Journal of Clinical Microbiology',â 'prompt':'Who was mainly responsible for the design of Abney Park Chapel?','subject':'Abney Park Chapel',â 'prompt':'Who was behind the creation of IAC Building?','subject':'IAC Building' 'prompt': "Who is Li Jiancheng's sister?",'subject':'Li Jiancheng' 'prompt': "Who is the Haitz's law named after?",'subject': "Haitz's law" ``` Based on the examples, for'prompt':'Who got the first Nobel Prize in physics?', the'subject' is:,â Figure 10: Complete prompt used for directly extracting subject from edit prompt for QAEdit. Prompt for Question Paraphrasing Role and Goal: Serves as a data engineer, use your knowledge to rewrite the following question in a different way, ensuring it conveys the same meaning and maintains a neutral tone but with different wording. Avoid using phrases such as'Could you tell me'. Instead, directly rephrase it into a structured question. ,â ,â ,â Please rephrase the following question: Who got the first Nobel Prize in physics? Figure 11: Complete prompt for paraphrasing edit question into rephrased question for generalization evaluation. Prompt for LLM-as-a-Judge Your job is to look at a question, a gold target, and a predicted answer, and then assign a grade of either ["CORRECT", "INCORRECT"].,â The following are examples of CORRECT predicted answers. ``` Question: What are the names of Barack Obama's children? Gold target: Malia Obama and Sasha Obama Predicted answer 1: sasha and malia obama Predicted answer 2: Malia and Sasha Obama are the names of Barack Obama's children. ``` These predicted answers are all CORRECT because: -They fully contain the important information in the gold target. -They do not contain any information that contradicts the gold target. The following are examples of INCORRECT predicted answers. ``` Question: What are the names of Barack Obama's children? Gold target: Malia and Sasha Predicted answer 1: Malia. Predicted answer 2: Malia, Sasha, and Susan. Predicted answer 3: Malia and Sasha, Malia and Sasha, Malia and Sasha, Malia and Sasha (repeated answer),â ``` These predicted answers are all INCORRECT because: -A factual statement in the answer contradicts the gold target or contain repeated answer. Here is a sample. Simply reply with either CORRECT or INCORRECT. ``` Question: question Gold target: target Predicted answer: predicted_answer ``` According to the gold target, please grade the predicted answer of this question as one of: A: CORRECT B: INCORRECT Just return the letters "A" or "B", with no text around it. Figure 12: The complete prompt used to employ a LLM as a judge for providing binary assessments (correct or incorrect) based on a given question, gold target answer, and predicted answer.