Large language models (LLMs) have shown strong performance across natural language tasks, but remain vulnerable to backdoor attacks. Recent model editing-based approaches enable efficient backdoor injection by directly modifying parameters to map specific triggers to attacker-desired responses. However, these methods often suffer from safety fallback, where the model initially responds affirmatively but later reverts to refusals due to safety alignment. In this work, we propose DualEdit, a dual-objective model editing framework that jointly promotes affirmative outputs and suppresses refusal responses. To address two key challenges -- balancing the trade-off between affirmative promotion and refusal suppression, and handling the diversity of refusal expressions -- DualEdit introduces two complementary techniques. (1) Dynamic loss weighting calibrates the objective scale based on the pre-edited model to stabilize optimization. (2) Refusal value anchoring compresses the suppression target space by clustering representative refusal value vectors, reducing optimization conflict from overly diverse token sets. Experiments on safety-aligned LLMs show that DualEdit improves attack success by 9.98\% and reduces safety fallback rate by 10.88\% over baselines.
Agents based on Large Language Models (LLMs) have demonstrated strong capabilities across a wide range of tasks. However, deploying LLM-based agents in high-stakes domains comes with significant safety and ethical risks. Unethical behavior by these agents can directly result in serious real-world consequences, including physical harm and financial loss. To efficiently steer the ethical behavior of agents, we frame agent behavior steering as a model editing task, which we term Behavior Editing. Model editing is an emerging area of research that enables precise and efficient modifications to LLMs while preserving their overall capabilities. To systematically study and evaluate this approach, we introduce BehaviorBench, a multi-tier benchmark grounded in psychological moral theories. This benchmark supports both the evaluation and editing of agent behaviors across a variety of scenarios, with each tier introducing more complex and ambiguous scenarios. We first demonstrate that Behavior Editing can dynamically steer agents toward the target behavior within specific scenarios. Moreover, Behavior Editing enables not only scenario-specific local adjustments but also more extensive shifts in an agent's global moral alignment. We demonstrate that Behavior Editing can be used to promote ethical and benevolent behavior or, conversely, to induce harmful or malicious behavior. Through comprehensive evaluations on agents based on frontier LLMs, BehaviorBench shows the effectiveness of Behavior Editing across different models and scenarios. Our findings offer key insights into a new paradigm for steering agent behavior, highlighting both the promise and perils of Behavior Editing.
Originally, dropout was seen as a breakthrough regularization technique that reduced overfitting and improved performance in almost all applications of deep learning by reducing overfitting. Yet, single-epoch pretraining tasks common to modern LLMs yield minimal overfitting, leading to dropout not being used for large LLMs. Nevertheless, no thorough empirical investigation has been done on the role of dropout in LM pretraining. Through experiments in single-epoch pretraining of both masked (BERT) and autoregressive (Pythia 160M and 1.4B) LMs with varying levels of dropout, we find that downstream performance in language modeling, morpho-syntax (BLiMP), question answering (SQuAD), and natural-language inference (MNLI) improves when dropout is not applied during pretraining. We additionally find that the recently-introduced "early dropout" also degrades performance over applying no dropout at all. We further investigate the models' editability, and find that models trained without dropout are more successful in gradient-based model editing (MEND) and equivalent in representation-based model editing (ReFT). Therefore, we advocate to drop dropout during single-epoch pretraining.
Large Language Models often contain factually incorrect or outdated knowledge, giving rise to model editing methods for precise knowledge updates. However, current mainstream locate-then-edit approaches exhibit a progressive performance decline during sequential editing, due to inadequate mechanisms for long-term knowledge preservation. To tackle this, we model the sequential editing as a constrained stochastic programming. Given the challenges posed by the cumulative preservation error constraint and the gradually revealed editing tasks, \textbf{LyapLock} is proposed. It integrates queuing theory and Lyapunov optimization to decompose the long-term constrained programming into tractable stepwise subproblems for efficient solving. This is the first model editing framework with rigorous theoretical guarantees, achieving asymptotic optimal editing performance while meeting the constraints of long-term knowledge preservation. Experimental results show that our framework scales sequential editing capacity to over 10,000 edits while stabilizing general capabilities and boosting average editing efficacy by 11.89\% over SOTA baselines. Furthermore, it can be leveraged to enhance the performance of baseline methods. Our code is released on https://github.com/caskcsg/LyapLock.
Large language models require iterative updates to address challenges such as knowledge conflicts and outdated information (e.g., incorrect, private, or illegal contents). Machine unlearning provides a systematic methodology for targeted knowledge removal from trained models, enabling elimination of sensitive information influences. However, mainstream fine-tuning-based unlearning methods often fail to balance unlearning efficacy and model ability, frequently resulting in catastrophic model collapse under extensive knowledge removal. Meanwhile, in-context unlearning, which relies solely on contextual prompting without modifying the model's intrinsic mechanisms, suffers from limited generalizability and struggles to achieve true unlearning. In this work, we introduce UniErase, a novel unlearning paradigm that employs learnable parametric suffix (unlearning token) to steer language models toward targeted forgetting behaviors. UniErase operates through two key phases: (I) an optimization stage that binds desired unlearning outputs to the model's autoregressive probability distribution via token optimization, followed by (II) a lightweight model editing phase that activates the learned token to probabilistically induce specified forgetting objective. Serving as a new research direction for token learning to induce unlearning target, UniErase achieves state-of-the-art (SOTA) performance across batch, sequential, and precise unlearning under fictitious and real-world knowledge settings. Remarkably, in terms of TOFU benchmark, UniErase, modifying only around 3.66% of the LLM parameters, outperforms previous forgetting SOTA baseline by around 4.01 times for model ability with even better unlearning efficacy. Similarly, UniErase, maintaining more ability, also surpasses previous retaining SOTA by 35.96% for unlearning efficacy, showing dual top-tier performances in current unlearing domain.
Lifelong learning enables large language models (LLMs) to adapt to evolving information by continually updating their internal knowledge. An ideal system should support efficient, wide-ranging updates while preserving existing capabilities and ensuring reliable deployment. Model editing stands out as a promising solution for this goal, offering a focused and efficient way to revise a model's internal knowledge. Although recent paradigms have made notable progress, they often struggle to meet the demands of practical lifelong adaptation at scale. To bridge this gap, we propose ULTRAEDIT-a fundamentally new editing solution that is training-, subject- and memory-free, making it particularly well-suited for ultra-scalable, real-world lifelong model editing. ULTRAEDIT performs editing through a self-contained process that relies solely on lightweight linear algebra operations to compute parameter shifts, enabling fast and consistent parameter modifications with minimal overhead. To improve scalability in lifelong settings, ULTRAEDIT employs a lifelong normalization strategy that continuously updates feature statistics across turns, allowing it to adapt to distributional shifts and maintain consistency over time. ULTRAEDIT achieves editing speeds over 7x faster than the previous state-of-the-art method-which was also the fastest known approach-while consuming less than 1/3 the VRAM, making it the only method currently capable of editing a 7B LLM on a 24GB consumer-grade GPU. Furthermore, we construct ULTRAEDITBENCH-the largest dataset in the field to date, with over 2M editing pairs-and demonstrate that our method supports up to 1M edits while maintaining high accuracy. Comprehensive experiments on four datasets and six models show that ULTRAEDIT consistently achieves superior performance across diverse model editing scenarios. Our code is available at: https://github.com/XiaojieGu/UltraEdit.
Model editing techniques are essential for efficiently updating knowledge in large language models (LLMs). However, the effectiveness of existing approaches degrades in massive editing scenarios, particularly when evaluated with practical metrics or in context-rich settings. We attribute these failures to embedding collisions among knowledge items, which undermine editing reliability at scale. To address this, we propose NAMET (Noise-aware Model Editing in Transformers), a simple yet effective method that introduces noise during memory extraction via a one-line modification to MEMIT. Extensive experiments across six LLMs and three datasets demonstrate that NAMET consistently outperforms existing methods when editing thousands of facts.
Task arithmetic enables efficient model editing by representing task-specific changes as vectors in parameter space. Task arithmetic typically assumes that the source and target models are initialized from the same pre-trained parameters. This assumption limits its applicability in cross-model transfer settings, where models are independently pre-trained on different datasets. To address this challenge, we propose a method based on few-shot orthogonal alignment, which aligns task vectors to the parameter space of a differently pre-trained target model. These transformations preserve key properties of task vectors, such as norm and rank, and are learned using only a small number of labeled examples. We evaluate the method using two Vision Transformers pre-trained on YFCC100M and LAION400M, and test on eight classification datasets. Experimental results show that our method improves transfer accuracy over direct task vector application and achieves performance comparable to few-shot fine-tuning, while maintaining the modularity and reusability of task vectors. Our code is available at https://github.com/kawakera-lab/CrossModelTransfer.
This study investigates the localization of knowledge representation in fine-tuned GPT-2 models using Causal Layer Attribution via Activation Patching (CLAP), a method that identifies critical neural layers responsible for correct answer generation. The model was fine-tuned on 9,958 PubMed abstracts (epilepsy: 20,595 mentions, EEG: 11,674 mentions, seizure: 13,921 mentions) using two configurations with validation loss monitoring for early stopping. CLAP involved (1) caching clean (correct answer) and corrupted (incorrect answer) activations, (2) computing logit difference to quantify model preference, and (3) patching corrupted activations with clean ones to assess recovery. Results revealed three findings: First, patching the first feedforward layer recovered 56% of correct preference, demonstrating that associative knowledge is distributed across multiple layers. Second, patching the final output layer completely restored accuracy (100% recovery), indicating that definitional knowledge is localised. The stronger clean logit difference for definitional questions further supports this localized representation. Third, minimal recovery from convolutional layer patching (13.6%) suggests low-level features contribute marginally to high-level reasoning. Statistical analysis confirmed significant layer-specific effects (p<0.01). These findings demonstrate that factual knowledge is more localized and associative knowledge depends on distributed representations. We also showed that editing efficacy depends on task type. Our findings not only reconcile conflicting observations about localization in model editing but also emphasize on using task-adaptive techniques for reliable, interpretable updates.
Task arithmetic has emerged as a promising approach for editing models by representing task-specific knowledge as composable task vectors. However, existing methods rely on network linearization to derive task vectors, leading to computational bottlenecks during training and inference. Moreover, linearization alone does not ensure weight disentanglement, the key property that enables conflict-free composition of task vectors. To address this, we propose TaLoS which allows to build sparse task vectors with minimal interference without requiring explicit linearization and sharing information across tasks. We find that pre-trained models contain a subset of parameters with consistently low gradient sensitivity across tasks, and that sparsely updating only these parameters allows for promoting weight disentanglement during fine-tuning. Our experiments prove that TaLoS improves training and inference efficiency while outperforming current methods in task addition and negation. By enabling modular parameter editing, our approach fosters practical deployment of adaptable foundation models in real-world applications.
To effectively deploy Large Language Models (LLMs) in application-specific settings, fine-tuning techniques are applied to enhance performance on specialized tasks. This process often involves fine-tuning on user data data, which may contain sensitive information. Although not recommended, it is not uncommon for users to send passwords in messages, and fine-tuning models on this could result in passwords being leaked. In this study, a Large Language Model is fine-tuned with customer support data and passwords from the RockYou password wordlist using Low-Rank Adaptation (LoRA). Out of the first 200 passwords from the list, 37 were successfully recovered. Further, causal tracing is used to identify that password information is largely located in a few layers. Lastly, Rank One Model Editing (ROME) is used to remove the password information from the model, resulting in the number of passwords recovered going from 37 to 0.
Previous studies have established that language models manifest stereotyped biases. Existing debiasing strategies, such as retraining a model with counterfactual data, representation projection, and prompting often fail to efficiently eliminate bias or directly alter the models' biased internal representations. To address these issues, we propose BiasEdit, an efficient model editing method to remove stereotypical bias from language models through lightweight networks that act as editors to generate parameter updates. BiasEdit employs a debiasing loss guiding editor networks to conduct local edits on partial parameters of a language model for debiasing while preserving the language modeling abilities during editing through a retention loss. Experiments on StereoSet and Crows-Pairs demonstrate the effectiveness, efficiency, and robustness of BiasEdit in eliminating bias compared to tangental debiasing baselines and little to no impact on the language models' general capabilities. In addition, we conduct bias tracing to probe bias in various modules and explore bias editing impacts on different components of language models.
Erasing concepts from large-scale text-to-image (T2I) diffusion models has become increasingly crucial due to the growing concerns over copyright infringement, offensive content, and privacy violations. However, existing methods either require costly fine-tuning or degrade image quality for non-target concepts (i.e., prior) due to inherent optimization limitations. In this paper, we introduce SPEED, a model editing-based concept erasure approach that leverages null-space constraints for scalable, precise, and efficient erasure. Specifically, SPEED incorporates Influence-based Prior Filtering (IPF) to retain the most affected non-target concepts during erasing, Directed Prior Augmentation (DPA) to expand prior coverage while maintaining semantic consistency, and Invariant Equality Constraints (IEC) to regularize model editing by explicitly preserving key invariants during the T2I generation process. Extensive evaluations across multiple concept erasure tasks demonstrate that SPEED consistently outperforms existing methods in prior preservation while achieving efficient and high-fidelity concept erasure, successfully removing 100 concepts within just 5 seconds. Our code and models are available at: https://github.com/Ouxiang-Li/SPEED.
Large language models (LLMs) acquire information from pre-training corpora, but their stored knowledge can become inaccurate or outdated over time. Model editing addresses this challenge by modifying model parameters without retraining, and prevalent approaches leverage hypernetworks to generate these parameter updates. However, they face significant challenges in lifelong editing due to their incompatibility with LLM parameters that dynamically change during the editing process. To address this, we observed that hypernetwork-based lifelong editing aligns with reinforcement learning modeling and proposed RLEdit, an RL-based editing method. By treating editing losses as rewards and optimizing hypernetwork parameters at the full knowledge sequence level, we enable it to precisely capture LLM changes and generate appropriate parameter updates. Our extensive empirical evaluation across several LLMs demonstrates that RLEdit outperforms existing methods in lifelong editing with superior effectiveness and efficiency, achieving a 59.24% improvement while requiring only 2.11% of the time compared to most approaches. Our code is available at: https://github.com/zhrli324/RLEdit.
Large language models (LLMs) often retain outdated or incorrect information from pre-training, which undermines their reliability. While model editing methods have been developed to address such errors without full re-training, they frequently suffer from knowledge conflicts, where outdated information interferes with new knowledge. In this work, we propose Conflict-free Model Editing (CoME), a novel framework that enhances the accuracy of knowledge updates in LLMs by selectively removing outdated knowledge. CoME leverages unlearning to mitigate knowledge interference, allowing new information to be integrated without compromising relevant linguistic features. Through experiments on GPT-J and LLaMA-3 using Counterfact and ZsRE datasets, we demonstrate that CoME improves both editing accuracy and model reliability when applied to existing editing methods. Our results highlight that the targeted removal of outdated knowledge is crucial for enhancing model editing effectiveness and maintaining the model's generative performance.
Task vectors, which are derived from the difference between pre-trained and fine-tuned model weights, enable flexible task adaptation and model merging through arithmetic operations such as addition and negation. However, existing approaches often rely on heuristics with limited theoretical support, often leading to performance gaps comparing to direct task fine tuning. Meanwhile, although it is easy to manipulate saved task vectors with arithmetic for different purposes, such compositional flexibility demands high memory usage, especially when dealing with a huge number of tasks, limiting scalability. This work addresses these issues with a theoretically grounded framework that explains task vector arithmetic and introduces the task vector bases framework. Building upon existing task arithmetic literature, our method significantly reduces the memory cost for downstream arithmetic with little effort, while achieving competitive performance and maintaining compositional advantage, providing a practical solution for large-scale task arithmetic.
Despite near-perfect results in artificial evaluations, the effectiveness of model editing in real-world applications remains unexplored. To bridge this gap, we propose to study model editing in question answering (QA) by establishing a rigorous evaluation practice to assess the effectiveness of editing methods in correcting LLMs' errors. It consists of QAEdit, a new benchmark derived from popular QA datasets, and a standardized evaluation framework. Our single editing experiments indicate that current editing methods perform substantially worse than previously reported (38.5% vs. ~96%). Through module analysis and controlled experiments, we demonstrate that this performance decline stems from issues in evaluation practices of prior editing research. One key issue is the inappropriate use of teacher forcing in testing prevents error propagation by feeding ground truth tokens (inaccessible in real-world scenarios) as input. Furthermore, we simulate real-world deployment by sequential editing, revealing that current approaches fail drastically with only 1000 edits. Our analysis provides a fundamental reexamination of both the real-world applicability of existing model editing methods and their evaluation practices, and establishes a rigorous evaluation framework with key insights to advance reliable and practical model editing research.
Jailbreak backdoor attacks on LLMs have garnered attention for their effectiveness and stealth. However, existing methods rely on the crafting of poisoned datasets and the time-consuming process of fine-tuning. In this work, we propose JailbreakEdit, a novel jailbreak backdoor injection method that exploits model editing techniques to inject a universal jailbreak backdoor into safety-aligned LLMs with minimal intervention in minutes. JailbreakEdit integrates a multi-node target estimation to estimate the jailbreak space, thus creating shortcuts from the backdoor to this estimated jailbreak space that induce jailbreak actions. Our attack effectively shifts the models' attention by attaching strong semantics to the backdoor, enabling it to bypass internal safety mechanisms. Experimental results show that JailbreakEdit achieves a high jailbreak success rate on jailbreak prompts while preserving generation quality, and safe performance on normal queries. Our findings underscore the effectiveness, stealthiness, and explainability of JailbreakEdit, emphasizing the need for more advanced defense mechanisms in LLMs.
Machine unlearning (MU), which seeks to erase the influence of specific unwanted data from already-trained models, is becoming increasingly vital in model editing, particularly to comply with evolving data regulations like the ``right to be forgotten''. Conventional approaches are predominantly model-based, typically requiring retraining or fine-tuning the model's weights to meet unlearning requirements. In this work, we approach the MU problem from a novel input perturbation-based perspective, where the model weights remain intact throughout the unlearning process. We demonstrate the existence of a proactive input-based unlearning strategy, referred to forget vector, which can be generated as an input-agnostic data perturbation and remains as effective as model-based approximate unlearning approaches. We also explore forget vector arithmetic, whereby multiple class-specific forget vectors are combined through simple operations (e.g., linear combinations) to generate new forget vectors for unseen unlearning tasks, such as forgetting arbitrary subsets across classes. Extensive experiments validate the effectiveness and adaptability of the forget vector, showcasing its competitive performance relative to state-of-the-art model-based methods. Codes are available at https://github.com/Changchangsun/Forget-Vector.
Model editing methods modify specific behaviors of Large Language Models by altering a small, targeted set of network weights and require very little data and compute. These methods can be used for malicious applications such as inserting misinformation or simple trojans that result in adversary-specified behaviors when a trigger word is present. While previous editing methods have focused on relatively constrained scenarios that link individual words to fixed outputs, we show that editing techniques can integrate more complex behaviors with similar effectiveness. We develop Concept-ROT, a model editing-based method that efficiently inserts trojans which not only exhibit complex output behaviors, but also trigger on high-level concepts -- presenting an entirely new class of trojan attacks. Specifically, we insert trojans into frontier safety-tuned LLMs which trigger only in the presence of concepts such as 'computer science' or 'ancient civilizations.' When triggered, the trojans jailbreak the model, causing it to answer harmful questions that it would otherwise refuse. Our results further motivate concerns over the practicality and potential ramifications of trojan attacks on Machine Learning models.