Methods for knowledge editing and unlearning in large language models seek to edit or remove undesirable knowledge or capabilities without compromising general language modeling performance. This work investigates how mechanistic interpretability -- which, in part, aims to identify model components (circuits) associated to specific interpretable mechanisms that make up a model capability -- can improve the precision and effectiveness of editing and unlearning. We find a stark difference in unlearning and edit robustness when training components localized by different methods. We highlight an important distinction between methods that localize components based primarily on preserving outputs, and those finding high level mechanisms with predictable intermediate states. In particular, localizing edits/unlearning to components associated with the lookup-table mechanism for factual recall 1) leads to more robust edits/unlearning across different input/output formats, and 2) resists attempts to relearn the unwanted information, while also reducing unintended side effects compared to baselines, on both a sports facts dataset and the CounterFact dataset across multiple models. We also find that certain localized edits disrupt the latent knowledge in the model more than any other baselines, making unlearning more robust to various attacks.
In-context knowledge editing (IKE) enables efficient modification of large language model (LLM) outputs without parameter changes and at zero-cost. However, it can be misused to manipulate responses opaquely, e.g., insert misinformation or offensive content. Such malicious interventions could be incorporated into high-level wrapped APIs where the final input prompt is not shown to end-users. To address this issue, we investigate the detection and reversal of IKE-edits. First, we demonstrate that IKE-edits can be detected with high accuracy (F1 > 80\%) using only the top-10 output probabilities of the next token, even in a black-box setting, e.g. proprietary LLMs with limited output information. Further, we introduce the novel task of reversing IKE-edits using specially tuned reversal tokens. We explore using both continuous and discrete reversal tokens, achieving over 80\% accuracy in recovering original, unedited outputs across multiple LLMs. Our continuous reversal tokens prove particularly effective, with minimal impact on unedited prompts. Through analysis of output distributions, attention patterns, and token rankings, we provide insights into IKE's effects on LLMs and how reversal tokens mitigate them. This work represents a significant step towards enhancing LLM resilience against potential misuse of in-context editing, improving their transparency and trustworthiness.
Large language models (LLMs) acquire knowledge during pre-training, but over time, this knowledge may become incorrect or outdated, necessitating updates after training. Knowledge editing techniques address this issue without the need for costly re-training. However, most existing methods are designed for single edits, and as the number of edits increases, they often cause a decline in the model's overall performance, posing significant challenges for sequential editing. To overcome this, we propose Orthogonal Subspace Editing, O-Edit. This algorithm orthogonalizes the direction of each knowledge update, minimizing interference between successive updates and reducing the impact of new updates on unrelated knowledge. Our approach does not require replaying previously edited data and processes each edit knowledge on time. It can perform thousands of edits on mainstream LLMs, achieving an average performance improvement that is 4.2 times better than existing methods while effectively preserving the model's performance on downstream tasks, all with minimal additional parameter overhead.
Collaborative learning of large language models (LLMs) has emerged as a new paradigm for utilizing private data from different parties to guarantee efficiency and privacy. Meanwhile, Knowledge Editing (KE) for LLMs has also garnered increased attention due to its ability to manipulate the behaviors of LLMs explicitly, yet leaves the collaborative KE case (in which knowledge edits of multiple parties are aggregated in a privacy-preserving and continual manner) unexamined. To this end, this manuscript dives into the first investigation of collaborative KE, in which we start by carefully identifying the unique three challenges therein, including knowledge overlap, knowledge conflict, and knowledge forgetting. We then propose a non-destructive collaborative KE framework, COLLABEDIT, which employs a novel model merging mechanism to mimic the global KE behavior while preventing the severe performance drop. Extensive experiments on two canonical datasets demonstrate the superiority of COLLABEDIT compared to other destructive baselines, and results shed light on addressing three collaborative KE challenges and future applications. Our code is available at https://github.com/LINs-lab/CollabEdit.
Large language models (LLMs) have revolutionized knowledge storage and retrieval, but face challenges with conflicting and outdated information. Knowledge editing techniques have been proposed to address these issues, yet they struggle with robustness tests involving long contexts, paraphrased subjects, and continuous edits. This work investigates the cause of these failures in locate-and-edit methods, offering theoretical insights into their key-value modeling and deriving mathematical bounds for robust and specific edits, leading to a novel 'group discussion' conceptual model for locate-and-edit methods. Empirical analysis reveals that keys used by current methods fail to meet robustness and specificity requirements. To address this, we propose a Robust Edit Pathway (REP) that disentangles editing keys from LLMs' inner representations. Evaluations on LLaMA2-7B and Mistral-7B using the CounterFact dataset show that REP significantly improves robustness across various metrics, both in-domain and out-of-domain, with minimal trade-offs in success rate and locality. Our findings advance the development of reliable and flexible knowledge updating in LLMs.
Knowledge editing has been proposed as an effective method for updating and correcting the internal knowledge of Large Language Models (LLMs). However, existing editing methods often struggle with complex tasks, such as multi-hop reasoning. In this paper, we identify and investigate the phenomenon of Editing Overfit, where edited models assign disproportionately high probabilities to the edit target, hindering the generalization of new knowledge in complex scenarios. We attribute this issue to the current editing paradigm, which places excessive emphasis on the direct correspondence between the input prompt and the edit target for each edit sample. To further explore this issue, we introduce a new benchmark, EVOKE (EValuation of Editing Overfit in Knowledge Editing), along with fine-grained evaluation metrics. Through comprehensive experiments and analysis, we demonstrate that Editing Overfit is prevalent in current editing methods and that common overfitting mitigation strategies are of limited effectiveness in knowledge editing. To overcome this, inspired by LLMs' knowledge recall mechanisms, we propose a new plug-and-play strategy called Learn to Inference (LTI), which introduce a Multi-stage Inference Constraint module to guide the edited models in recalling new knowledge similarly to how unedited LLMs leverage knowledge through in-context learning. Extensive experimental results across a wide range of tasks validate the effectiveness of LTI in mitigating Editing Overfit.
In recent years, with the maturation of large language model (LLM) technology and the emergence of high-quality programming code datasets, researchers have become increasingly confident in addressing the challenges of program synthesis automatically. However, since most of the training samples for LLMs are unscreened, it is inevitable that LLMs' performance may not align with real-world scenarios, leading to the presence of social bias. To evaluate and quantify the gender bias in code LLMs, we propose a dataset named CodeGenBias (Gender Bias in the Code Generation) and an evaluation metric called FB-Score (Factual Bias Score) based on the actual gender distribution of correlative professions. With the help of CodeGenBias and FB-Score, we evaluate and analyze the gender bias in eight mainstream Code LLMs. Previous work has demonstrated that model editing methods that perform well in knowledge editing have the potential to mitigate social bias in LLMs. Therefore, we develop a model editing approach named MG-Editing (Multi-Granularity model Editing), which includes the locating and editing phases. Our model editing method MG-Editing can be applied at five different levels of model parameter granularity: full parameters level, layer level, module level, row level, and neuron level. Extensive experiments not only demonstrate that our MG-Editing can effectively mitigate the gender bias in code LLMs while maintaining their general code generation capabilities, but also showcase its excellent generalization. At the same time, the experimental results show that, considering both the gender bias of the model and its general code generation capability, MG-Editing is most effective when applied at the row and neuron levels of granularity.
The locate-then-edit paradigm has shown significant promise for knowledge editing (KE) in Large Language Models (LLMs). While previous methods perform well on single-hop fact recall tasks, they consistently struggle with multi-hop factual recall tasks involving newly edited knowledge. In this paper, leveraging tools in mechanistic interpretability, we first identify that in multi-hop tasks, LLMs tend to retrieve knowledge with implicit subject information from deeper MLP layers, unlike single-hop tasks, which rely on shallow layers. This distinction explains the poor performance of current methods in multi-hop queries, as they primarily focus on editing shallow layers with single-hop edit prompts, leaving deeper layers unchanged. To address this, we propose IFMET, a novel locate-then-edit KE approach designed to edit both shallow and deep MLP layers. Beyond single-hop editing prompts, IFMET further incorporates multi-hop editing prompts to locate and modify knowledge across different stages of reasoning. Experimental results demonstrate that IFMET significantly improves performance on multi-hop factual recall tasks, overcoming the limitations of previous locate-then-edit methods
Large language models (LLMs) have demonstrated remarkable capabilities, but they also pose risks related to the generation of toxic or harmful content. This work introduces Precision Knowledge Editing (PKE), an advanced technique that builds upon existing knowledge editing methods to more effectively identify and modify toxic parameter regions within LLMs. By leveraging neuron weight tracking and activation pathway tracing, PKE achieves finer granularity in toxic content management compared to previous methods like Detoxifying Instance Neuron Modification (DINM). Our experiments demonstrate that PKE significantly reduces the attack success rate (ASR) across various models, including Llama2-7b and Llama-3-8b-instruct, while maintaining overall model performance. Additionally, we also compared the performance of some closed-source models (gpt-4-0613 and Claude 3 Sonnet) in our experiments, and found that models adjusted using our method far outperformed the closed-source models in terms of safety. This research contributes to the ongoing efforts to make LLMs safer and more reliable for real-world applications.
The ripple effect poses a significant challenge in knowledge editing for large language models. Namely, when a single fact is edited, the model struggles to accurately update the related facts in a sequence, which is evaluated by multi-hop questions linked to a chain of related facts. Recent strategies have moved away from traditional parameter updates to more flexible, less computation-intensive methods, proven to be more effective in addressing the ripple effect. In-context learning (ICL) editing uses a simple demonstration `Imagine that + new fact` to guide LLMs, but struggles with complex multi-hop questions as the new fact alone fails to specify the chain of facts involved in such scenarios. Besides, memory-based editing maintains additional storage for all edits and related facts, requiring continuous updates to stay effective. As a result of these design limitations, the challenge remains, with the highest accuracy being only 33.8% on the MQuAKE-cf benchmarks for Vicuna-7B. To address this, we propose RippleCOT, a novel ICL editing approach integrating Chain-of-Thought (COT) reasoning. RippleCOT structures demonstrations as `newfact, question, thought, answer`, incorporating a thought component to identify and decompose the multi-hop logic within questions. This approach effectively guides the model through complex multi-hop questions with chains of related facts. Comprehensive experiments demonstrate that RippleCOT significantly outperforms the state-of-the-art on the ripple effect, achieving accuracy gains ranging from 7.8% to 87.1%.
Large language models (LLMs) often exhibit hallucinations due to incorrect or outdated knowledge. Hence, model editing methods have emerged to enable targeted knowledge updates. To achieve this, a prevailing paradigm is the locating-then-editing approach, which first locates influential parameters and then edits them by introducing a perturbation. While effective, current studies have demonstrated that this perturbation inevitably disrupt the originally preserved knowledge within LLMs, especially in sequential editing scenarios. To address this, we introduce AlphaEdit, a novel solution that projects perturbation onto the null space of the preserved knowledge before applying it to the parameters. We theoretically prove that this projection ensures the output of post-edited LLMs remains unchanged when queried about the preserved knowledge, thereby mitigating the issue of disruption. Extensive experiments on various LLMs, including LLaMA3, GPT2-XL, and GPT-J, show that AlphaEdit boosts the performance of most locating-then-editing methods by an average of 36.4% with a single line of additional code for projection solely. Our code is available at: https://github.com/jianghoucheng/AlphaEdit.
The swift advancement in Multimodal LLMs (MLLMs) also presents significant challenges for effective knowledge editing. Current methods, including intrinsic knowledge editing and external knowledge resorting, each possess strengths and weaknesses, struggling to balance the desired properties of reliability, generality, and locality when applied to MLLMs. In this paper, we propose UniKE, a novel multimodal editing method that establishes a unified perspective and paradigm for intrinsic knowledge editing and external knowledge resorting. Both types of knowledge are conceptualized as vectorized key-value memories, with the corresponding editing processes resembling the assimilation and accommodation phases of human cognition, conducted at the same semantic levels. Within such a unified framework, we further promote knowledge collaboration by disentangling the knowledge representations into the semantic and truthfulness spaces. Extensive experiments validate the effectiveness of our method, which ensures that the post-edit MLLM simultaneously maintains excellent reliability, generality, and locality. The code for UniKE is available at \url{https://github.com/beepkh/UniKE}.
Knowledge editing has emerged as an efficient technology for updating the knowledge of large language models (LLMs), attracting increasing attention in recent years. However, there is a lack of effective measures to prevent the malicious misuse of this technology, which could lead to harmful edits in LLMs. These malicious modifications could cause LLMs to generate toxic content, misleading users into inappropriate actions. In front of this risk, we introduce a new task, Knowledge Editing Type Identification (KETI), aimed at identifying different types of edits in LLMs, thereby providing timely alerts to users when encountering illicit edits. As part of this task, we propose KETIBench, which includes five types of harmful edits covering most popular toxic types, as well as one benign factual edit. We develop four classical classification models and three BERT-based models as baseline identifiers for both open-source and closed-source LLMs. Our experimental results, across 42 trials involving two models and three knowledge editing methods, demonstrate that all seven baseline identifiers achieve decent identification performance, highlighting the feasibility of identifying malicious edits in LLMs. Additional analyses reveal that the performance of the identifiers is independent of the reliability of the knowledge editing methods and exhibits cross-domain generalization, enabling the identification of edits from unknown sources. All data and code are available in https://github.com/xpq-tech/KETI. Warning: This paper contains examples of toxic text.
Training large language models (LLMs) from scratch is an expensive endeavor, particularly as world knowledge continually evolves. To maintain relevance and accuracy of LLMs, model editing has emerged as a pivotal research area. While these methods hold promise, they can also produce unintended side effects. Their underlying factors and causes remain largely unexplored. This paper delves into a critical factor-question type-by categorizing model editing questions. Our findings reveal that the extent of performance degradation varies significantly across different question types, providing new insights for experimental design in knowledge editing. Furthermore, we investigate whether insights from smaller models can be extrapolated to larger models. Our results indicate discrepancies in findings between models of different sizes, suggesting that insights from smaller models may not necessarily apply to larger models. Additionally, we examine the impact of batch size on side effects, discovering that increasing the batch size can mitigate performance drops.
During pre-training, the Text-to-Image (T2I) diffusion models encode factual knowledge into their parameters. These parameterized facts enable realistic image generation, but they may become obsolete over time, thereby misrepresenting the current state of the world. Knowledge editing techniques aim to update model knowledge in a targeted way. However, facing the dual challenges posed by inadequate editing datasets and unreliable evaluation criterion, the development of T2I knowledge editing encounter difficulties in effectively generalizing injected knowledge. In this work, we design a T2I knowledge editing framework by comprehensively spanning on three phases: First, we curate a dataset \textbf{CAKE}, comprising paraphrase and multi-object test, to enable more fine-grained assessment on knowledge generalization. Second, we propose a novel criterion, \textbf{adaptive CLIP threshold}, to effectively filter out false successful images under the current criterion and achieve reliable editing evaluation. Finally, we introduce \textbf{MPE}, a simple but effective approach for T2I knowledge editing. Instead of tuning parameters, MPE precisely recognizes and edits the outdated part of the conditioning text-prompt to accommodate the up-to-date knowledge. A straightforward implementation of MPE (Based on in-context learning) exhibits better overall performance than previous model editors. We hope these efforts can further promote faithful evaluation of T2I knowledge editing methods.
Large Language Models (LLMs) have demonstrated significant capabilities across numerous application domains. A key challenge is to keep these models updated with latest available information, which limits the true potential of these models for the end-applications. Although, there have been numerous attempts for LLMs Knowledge Editing (KE), i.e., to edit the LLMs prior knowledge and in turn test it via Multi-hop Question Answering (MQA), yet so far these studies are primarily focused on English language. To bridge this gap, in this paper we propose: Multi-hop Questioning Answering under Knowledge Editing for Arabic Language (MQA-KEAL). MQA-KEAL stores knowledge edits as structured knowledge units in the external memory. In order to solve multi-hop question, it first uses task-decomposition to decompose the question into smaller sub-problems. Later for each sub-problem, it iteratively queries the external memory and/or target LLM in order to generate the final response. In addition, we also contribute MQUAKE-AR (Arabic translation of English benchmark MQUAKE), as well as a new benchmark MQA-AEVAL for rigorous performance evaluation of MQA under KE for Arabic language. Experimentation evaluation reveals MQA-KEAL outperforms the baseline models by a significant margin.
The manipulation of the personality traits of large language models (LLMs) has emerged as a key area of research. Methods like prompt-based In-Context Knowledge Editing (IKE) and gradient-based Model Editor Networks (MEND) have been explored but show irregularity and variability; IKE depends on the prompt, leading to variability and sensitivity, while MEND yields inconsistent and gibberish outputs. To address this, we employed Opinion QA Based Parameter-Efficient Fine-Tuning (PEFT), specifically Quantized Low-Rank Adaptation (QLoRA), to manipulate the Big Five personality traits: Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. After PEFT, models such as Mistral-7B-Instruct and LLaMA-2-7B-chat began generating emojis, even though no emojis were present in the PEFT data. For instance, LLaMA-2-7B-chat generated emojis in 99.5% of extraversion-related test instances, while Mistral-7B-Instruct did so in 92.5% of openness-related test instances. ICL Explainability analysis indicated that the LLMs used emojis intentionally to express these traits. Mechanistic Interpretability analysis showed that this latent behaviour of LLMs could be traced to specific neurons that became activated or amplified after PEFT. This paper provides a number of novel contributions. First, introducing an Opinion QA dataset for PEFT-driven personality manipulation; second, developing metric models to benchmark LLM personality traits; third, demonstrating PEFT's superiority over IKE in personality manipulation; and finally, analysing and validating emoji usage through explainability methods such as Mechanistic Interpretability and In-context learning Explainability methods.
As the modern tool of choice for question answering, large language models (LLMs) are expected to deliver answers with up-to-date knowledge. To achieve such ideal question-answering systems, locating and then editing outdated knowledge in the natural language outputs is a general target of popular knowledge editing methods. However, this target is challenging, as both identifying which tokens to edit in the reasoning steps and ensuring the coherence of the revised reasoning chain are difficult tasks. We argue that these challenges stem from the unstructured nature of natural language outputs. To address the above challenges, we propose $\textbf{Stru}$ctural $\textbf{Edit}$ing ($\textbf{StruEdit}$), an improved baseline for knowledge editing. We first prompt LLMs to produce structured outputs consisting of reasoning triplets. Then, StruEdit removes any potentially outdated knowledge and efficiently refills the structured outputs with up-to-date information in a single step. Experimental results show that StruEdit consistently delivers the highest accuracy with lowest latency compared with other knowledge editing methods.
Knowledge representation has been a central aim of AI since its inception. Symbolic Knowledge Graphs (KGs) and neural Large Language Models (LLMs) can both represent knowledge. KGs provide highly accurate and explicit knowledge representation, but face scalability issue; while LLMs offer expansive coverage of knowledge, but incur significant training costs and struggle with precise and reliable knowledge manipulation. To this end, we introduce OneEdit, a neural-symbolic prototype system for collaborative knowledge editing using natural language, which facilitates easy-to-use knowledge management with KG and LLM. OneEdit consists of three modules: 1) The Interpreter serves for user interaction with natural language; 2) The Controller manages editing requests from various users, leveraging the KG with rollbacks to handle knowledge conflicts and prevent toxic knowledge attacks; 3) The Editor utilizes the knowledge from the Controller to edit KG and LLM. We conduct experiments on two new datasets with KGs which demonstrate that OneEdit can achieve superior performance.
Chinese, as a linguistic system rich in depth and complexity, is characterized by distinctive elements such as ancient poetry, proverbs, idioms, and other cultural constructs. However, current Large Language Models (LLMs) face limitations in these specialized domains, highlighting the need for the development of comprehensive datasets that can assess, continuously update, and progressively improve these culturally-grounded linguistic competencies through targeted training optimizations. To address this gap, we introduce CKnowEdit, the first-ever Chinese knowledge editing dataset designed to correct linguistic, factual, and logical errors in LLMs. We collect seven types of knowledge from a wide range of sources, including classical texts, idioms, and content from Baidu Tieba Ruozhiba, taking into account the unique polyphony, antithesis, and logical structures inherent in the Chinese language. By analyzing this dataset, we highlight the challenges current LLMs face in mastering Chinese. Furthermore, our evaluation of state-of-the-art knowledge editing techniques reveals opportunities to advance the correction of Chinese knowledge. Code and dataset are available at https://github.com/zjunlp/EasyEdit.