Recent advancements in graph unlearning models have enhanced model utility by preserving the node representation essentially invariant, while using gradient ascent on the forget set to achieve unlearning. However, this approach causes a drastic degradation in model utility during the unlearning process due to the rapid divergence speed of gradient ascent. In this paper, we introduce \textbf{INPO}, an \textbf{I}nfluence-aware \textbf{N}egative \textbf{P}reference \textbf{O}ptimization framework that focuses on slowing the divergence speed and improving the robustness of the model utility to the unlearning process. Specifically, we first analyze that NPO has slower divergence speed and theoretically propose that unlearning high-influence edges can reduce impact of unlearning. We design an influence-aware message function to amplify the influence of unlearned edges and mitigate the tight topological coupling between the forget set and the retain set. The influence of each edge is quickly estimated by a removal-based method. Additionally, we propose a topological entropy loss from the perspective of topology to avoid excessive information loss in the local structure during unlearning. Extensive experiments conducted on five real-world datasets demonstrate that INPO-based model achieves state-of-the-art performance on all forget quality metrics while maintaining the model's utility. Codes are available at \href{https://github.com/sh-qiangchen/INPO}{https://github.com/sh-qiangchen/INPO}.

Large language models trained on vast corpora inherently risk memorizing sensitive or harmful content, which may later resurface in their outputs. Prevailing unlearning methods generally rely on gradient ascent and its variants to lower the probability of specific target responses. However, we find that this strategy induces a critical side effect: probability mass is redistributed into high-likelihood regions, often corresponding to semantically related rephrasings of the targets. We refer to this as the squeezing effect, which explains why many methods yield merely spurious unlearning, a problem further obscured by automated metrics (e.g., ROUGE, truth ratio) that misreport actual success. To address this, we propose a bootstrapping (BS) framework that explicitly links the squeezing effect with the model's own high-confidence generations, namely its model beliefs. Since model beliefs inherently capture the very high-likelihood regions where probability mass is squeezed, incorporating them into the unlearning objective directly counters the squeezing effect. By jointly suppressing both target responses and model beliefs, BS-T (token) attenuates high-probability tokens, whereas BS-S (sequence) removes entire high-confidence generations, together achieving more thorough forgetting while preserving utility. Extensive experiments across diverse benchmarks with various model families confirm the effectiveness of our approach.

Machine unlearning offers a promising solution to privacy and safety concerns in large language models (LLMs) by selectively removing targeted knowledge while preserving utility. However, current methods are highly sensitive to downstream fine-tuning, which can quickly recover forgotten information-even from unrelated tasks. To address this, we introduce invariance into unlearning for the first time, inspired by invariant risk minimization (IRM). Building on this principle, we propose invariant LLM unlearning (ILU), a regularization-based framework that enhances robustness. Notably, ILU generalizes well to diverse fine-tuning tasks, even when trained using a single dataset. A task vector analysis is also provided to further elucidate the rationale behind ILU's effectiveness. Extensive experiments on the WMDP and MUSE benchmark, reveal that ILU significantly outperforms state-of-the-art unlearning methods, including negative preference optimization (NPO) and representation misdirection for unlearning (RMU). Notably, ILU achieves superior unlearning robustness across diverse downstream fine-tuning scenarios (e.g., math, paraphrase detection, and sentiment analysis) while preserving the fine-tuning performance.

Current LLM unlearning methods face a critical security vulnerability that undermines their fundamental purpose: while they appear to successfully remove sensitive or harmful knowledge, this ``forgotten" information remains precariously recoverable through relearning attacks. We identify that the root cause is that conventional methods optimizing the forgetting loss at individual data points will drive model parameters toward sharp minima in the loss landscape. In these unstable regions, even minimal parameter perturbations can drastically alter the model's behaviors. Consequently, relearning attacks exploit this vulnerability by using just a few fine-tuning samples to navigate the steep gradients surrounding these unstable regions, thereby rapidly recovering knowledge that was supposedly erased. This exposes a critical robustness gap between apparent unlearning and actual knowledge removal. To address this issue, we propose StableUN, a bi-level feedback-guided optimization framework that explicitly seeks more stable parameter regions via neighborhood-aware optimization. It integrates forgetting feedback, which uses adversarial perturbations to probe parameter neighborhoods, with remembering feedback to preserve model utility, aligning the two objectives through gradient projection. Experiments on WMDP and MUSE benchmarks demonstrate that our method is significantly more robust against both relearning and jailbreaking attacks while maintaining competitive utility performance.

Unlike prior approaches that treat alignment as a static or post-hoc property, NPO introduces a formalization of alignment loss that is measurable, supervisable, and reducible under structured feedback. In parallel, we propose meta-alignment as the fidelity of the monitoring process that governs retraining or override triggers, and show that it is formally reducible to primary alignment via threshold fidelity. Our implementation spans a scalable operational loop involving scenario scoring, threshold tuning, policy validation, and structured feedback ingestion, including "likes," overrides, and abstentions. We provide formal convergence results under stochastic feedback and show that both alignment loss and monitoring fidelity converge additively. Empirically, NPO demonstrates measurable value in hyperscale deployment settings. A simulation-based artifact and ablation studies further illustrate the theoretical principles in action. Together, NPO offers a compact, inspectable architecture for continual alignment monitoring, helping bridge theoretical alignment guarantees with practical reliability in dynamic environments.

Large language models (LLMs) should undergo rigorous audits to identify potential risks, such as copyright and privacy infringements. Once these risks emerge, timely updates are crucial to remove undesirable responses, ensuring legal and safe model usage. It has spurred recent research into LLM unlearning, focusing on erasing targeted undesirable knowledge without compromising the integrity of other, non-targeted responses. Existing studies have introduced various unlearning objectives to pursue LLM unlearning without necessitating complete retraining. However, each of these objectives has unique properties, and no unified framework is currently available to comprehend them thoroughly. To fill the gap, we propose a toolkit of the gradient effect (G-effect), quantifying the impacts of unlearning objectives on model performance from a gradient perspective. A notable advantage is its broad ability to detail the unlearning impacts from various aspects across instances, updating steps, and LLM layers. Accordingly, the G-effect offers new insights into identifying drawbacks of existing unlearning objectives, further motivating us to explore a series of new solutions for their mitigation and improvements. Finally, we outline promising directions that merit further studies, aiming at contributing to the community to advance this important field.

With the rapid advancement of large language models (LLMs), concerns about their privacy, safety, and trustworthiness, have become increasingly prominent (Liu et al., 2024d; Barez et al., 2025). However, retraining these models to eliminate the undesired data-model influence is often infeasible due to the significant computational and time costs involved. To address this challenge, LLM unlearning (Yao et al., 2024; Eldan & Russinovich, 2023; Maini et al., 2024; Liu et al., 2024b) has emerged as a post-pretraining strategy, which aims to mitigate the impact of undesirable data (e.g., sensitive, biased, unsafe, or illegal information) and suppress associated model capabilities, thereby preventing LLMs from generating harmful content while simultaneously preserving the model's utility post-unlearning. Despite the increasing importance of LLM unlearning, several recent studies (Łucki et al., 2024; Zhang et al., 2024e; Lynch et al., 2024; Hu et al., 2024; Deeb & Roger, 2024) have identified a critical issue: LLM unlearning often lacks robustness. Specifically, the susceptibility to quickly recovering'already-unlearned' knowledge post-unlearning is evident through so-called relearning attacks (Lynch et al., 2024; Hu et al., 2024). These attacks can effectively reverse the unlearning process by leveraging lightweight fine-tuning on the unlearned model using only a small number of data from the forget dataset.

Large Language Models (LLMs) often memorize sensitive, private, or copyrighted data during pre-training. LLM unlearning aims to eliminate the influence of undesirable data from the pre-trained model while preserving the model's utilities on other tasks. Several practical methods have recently been proposed for LLM unlearning, mostly based on gradient ascent (GA) on the loss of undesirable data. However, on certain unlearning tasks, these methods either fail to effectively unlearn the target data or suffer from catastrophic collapse -- a drastic degradation of the model's utilities. In this paper, we propose Negative Preference Optimization (NPO), a simple alignment-inspired method that could efficiently and effectively unlearn a target dataset. We theoretically show that the progression toward catastrophic collapse by minimizing the NPO loss is exponentially slower than GA. Through experiments on synthetic data and the benchmark TOFU dataset, we demonstrate that NPO-based methods achieve a better balance between unlearning the undesirable data and maintaining the model's utilities. We also observe that NPO-based methods generate more sensible outputs than GA-based methods, whose outputs are often gibberish. Remarkably, on TOFU, NPO-based methods are the first to achieve reasonable unlearning results in forgetting 50% (or more) of the training data, whereas existing methods already struggle with forgetting 10% of training data.

Out-of-distribution (OOD) detection is indispensable for safely deploying machine learning models in the wild. One of the key challenges is that models lack supervision signals from unknown data, and as a result, can produce overconfident predictions on OOD data. Recent work on outlier synthesis modeled the feature space as parametric Gaussian distribution, a strong and restrictive assumption that might not hold in reality. In this paper, we propose a novel framework, Non-Parametric Outlier Synthesis (NPOS), which generates artificial OOD training data and facilitates learning a reliable decision boundary between ID and OOD data. Importantly, our proposed synthesis approach does not make any distributional assumption on the ID embeddings, thereby offering strong flexibility and generality. We show that our synthesis approach can be mathematically interpreted as a rejection sampling framework. Extensive experiments show that NPOS can achieve superior OOD detection performance, outperforming the competitive rivals by a significant margin. Code is publicly available at https://github.com/deeplearning-wisc/npos.

A fundamental problem in multi-agent systems is resource allocation. Specifically, the problem of assigning a number of indivisible objects to agents with different preferences has been widely studied not only in multi-agent systems, but also in economics [21] and theoretical computer science [12]. The focus of our work is the special case of allocating n objects to n agents (so each agent is matched to a single object), which models many real-world applications. For example, imagine allocating office spaces to faculty members in a new building. Instead of asking each faculty member to report a full preference ranking over the available offices, the department head may ask faculty members to reveal their top choices, and then if need be, he may ask individual faculty members to reveal their next best choices, and so on. The goal of the department head is to learn a matching that satisfies some form of "economic efficiency" while asking as few queries as possible.