NeurIPS 2025 Papers

The ability to selectively remove knowledge from LLMs is highly desirable. However, existing methods often struggle with balancing unlearning efficacy and retain model utility, and lack controllability at inference time to emulate base model behavior as if it had never seen the unlearned data. In this paper, we propose LUNAR, a novel unlearning method grounded in the Linear Representation Hypothesis and operates by redirecting the representations of unlearned data to activation regions that expresses its inability to answer. We show that contrastive features are not a prerequisite for effective activation redirection, and LUNAR achieves state-of-the-art unlearning performance and superior controllability. Specifically, LUNAR achieves between 2.9x and 11.7x improvement in the combined unlearning efficacy and model utility score (Deviation Score) across various base models and generates coherent, contextually appropriate responses post-unlearning. Moreover, LUNAR effectively reduces parameter updates to a single down-projection matrix, a novel design that significantly enhances efficiency by 20x and robustness. Finally, we demonstrate that LUNAR is robust to white-box adversarial attacks and versatile in real-world scenarios, including handling sequential unlearning requests.

Harmful fine-tuning poses critical safety risks to fine-tuning-as-a-service for large language models. Existing defense strategies preemptively build robustness via attack simulation but suffer from fundamental limitations: (i) the infeasibility of extending attack simulations beyond bounded threat models due to the inherent difficulty of anticipating unknown attacks, and (ii) limited adaptability to varying attack settings, as simulation fails to capture their variability and complexity. To address these challenges, we propose Bayesian Data Scheduler (BDS), an adaptive tuning-stage defense strategy with no need for attack simulation. BDS formulates harmful fine-tuning defense as a Bayesian inference problem, learning the posterior distribution of each data point's safety attribute, conditioned on the fine-tuning and alignment datasets. The fine-tuning process is then constrained by weighting data with their safety attributes sampled from the posterior, thus mitigating the influence of harmful data. By leveraging the post hoc nature of Bayesian inference, the posterior is conditioned on the fine-tuning dataset, enabling BDS to tailor its defense to the specific dataset, thereby achieving adaptive defense. Furthermore, we introduce a neural scheduler based on amortized Bayesian learning, enabling efficient transfer to new data without retraining. Comprehensive results across diverse attack and defense settings demonstrate the state-of-the-art performance of our approach. Code is available at https://github.com/Egg-Hu/Bayesian-Data-Scheduler.

Harmful fine-tuning attack introduces significant security risks to the fine-tuning services. Main-stream defenses aim to vaccinate the model such that the later harmful fine-tuning attack is less effective. However, our evaluation results show that such defenses are fragile-- with a few fine-tuning steps, the model still can learn the harmful knowledge. To this end, we do further experiment and find that an embarrassingly simple solution-- adding purely random perturbations to the fine-tuned model, can recover the model from harmful behaviors, though it leads to a degradation in the model’s fine-tuning performance. To address the degradation of fine-tuning performance, we further propose \methodname, which optimizes an adaptive perturbation that will be applied to the model after fine-tuning. \methodname maintains model's safety alignment performance without compromising downstream fine-tuning performance. Comprehensive experiments are conducted on different harmful ratios, fine-tuning tasks and mainstream LLMs, where the average harmful scores are reduced by up-to 21.2%, while maintaining fine-tuning performance. As a by-product, we analyze the adaptive perturbation and show that different layers in various LLMs have distinct safety coefficients. Source code available at https://github.com/w-yibo/Panacea.

We study the reinforcement learning (RL) problem in a constrained Markov decision process (CMDP), where an agent explores the environment to maximize the expected cumulative reward while satisfying a single constraint on the expected total utility value in every episode. While this problem is well understood in the tabular setting, theoretical results for function approximation remain scarce. This paper closes the gap by proposing an RL algorithm for linear CMDPs that achieves $\widetilde{\mathcal{O}}(\sqrt{K})$ regret with an episode-wise zero-violation guarantee. Furthermore, our method is computationally efficient, scaling polynomially with problem-dependent parameters while remaining independent of the state space size. Our results significantly improve upon recent linear CMDP algorithms, which either violate the constraint or incur exponential computational costs.

Training data attribution (TDA) is concerned with understanding model behavior in terms of the training data. This paper draws attention to the common setting where one has access only to the final trained model, and not the training algorithm or intermediate information from training. We reframe the problem in this "final-model-only" setting as one of measuring sensitivity of the model to training instances. To operationalize this reframing, we propose *further training*, with appropriate adjustment and averaging, as a gold standard method to measure sensitivity. We then unify existing gradient-based methods for TDA by showing that they all approximate the further training gold standard in different ways. We investigate empirically the quality of these gradient-based approximations to further training, for tabular, image, and text datasets and models. We find that the approximation quality of first-order methods is sometimes high but decays with the amount of further training. In contrast, the approximations given by influence function methods are more stable but surprisingly lower in quality.

This paper proposes a new algorithm for learning gradient boosted decision trees while ensuring the existence of recourse actions. Algorithmic recourse aims to provide a recourse action for altering the undesired prediction result given by a model. While existing studies often focus on extracting valid and executable actions from a given learned model, such reasonable actions do not always exist for models optimized solely for predictive accuracy. To address this issue, recent studies proposed a framework for learning a model while guaranteeing the existence of reasonable actions with high probability. However, these methods can not be applied to gradient boosted decision trees, which are renowned as one of the most popular models for tabular datasets. We propose an efficient gradient boosting algorithm that takes recourse guarantee into account, while maintaining the same time complexity as the standard ones. We also propose a post-processing method for refining a learned model under the constraint of a recourse guarantee and provide a PAC-style analysis of the refined model. Experimental results demonstrated that our method successfully provided reasonable actions to more instances than the baselines without significantly degrading accuracy and computational efficiency.

Data attribution methods, which quantify the influence of individual training data points on a machine learning model, have gained increasing popularity in data-centric applications in modern AI. Despite a recent surge of new methods developed in this space, the impact of hyperparameter tuning in these methods remains under-explored. In this work, we present the first large-scale empirical study to understand the hyperparameter sensitivity of common data attribution methods. Our results show that most methods are indeed sensitive to certain key hyperparameters. However, unlike typical machine learning algorithms---whose hyperparameters can be tuned using computationally-cheap validation metrics---evaluating data attribution performance often requires retraining models on subsets of training data, making such metrics prohibitively costly for hyperparameter tuning. This poses a critical open challenge for the practical application of data attribution methods. To address this challenge, we advocate for better theoretical understandings of hyperparameter behavior to inform efficient tuning strategies. As a case study, we provide a theoretical analysis of the regularization term that is critical in many variants of influence function methods. Building on this analysis, we propose a lightweight procedure for selecting the regularization value without model retraining, and validate its effectiveness across a range of standard data attribution benchmarks. Overall, our study identifies a fundamental yet overlooked challenge in the practical application of data attribution, and highlights the importance of careful discussion on hyperparameter selection in future method development.

Large language models (LLMs) can sometimes report the strategies they actually use to solve tasks, yet at other times seem unable to recognize those strategies that govern their behavior. This suggests a limited degree of metacognition --- the capacity to monitor one's own cognitive processes for subsequent reporting and self-control. Metacognition enhances LLMs' capabilities in solving complex tasks but also raises safety concerns, as models may obfuscate their internal processes to evade neural-activation-based oversight (e.g., safety detector). Given society's increased reliance on these models, it is critical that we understand their metacognitive abilities. To address this, we introduce a neuroscience-inspired \emph{neurofeedback} paradigm that uses in-context learning to quantify metacognitive abilities of LLMs to \textit{report} and \textit{control} their activation patterns. We demonstrate that their abilities depend on several factors: the number of in-context examples provided, the semantic interpretability of the neural activation direction (to be reported/controlled), and the variance explained by that direction. These directions span a ``metacognitive space'' with dimensionality much lower than the model's neural space, suggesting LLMs can monitor only a small subset of their neural activations. Our paradigm provides empirical evidence to quantify metacognition in LLMs, with significant implications for AI safety (e.g., adversarial attack and defense).

A panel satisfies *descriptive representation* when its composition reflects the population. We examine the role of descriptive representation in collective decision making through an optimization lens, asking whether representative panels make decisions that maximize social welfare for the underlying population. Our main results suggest that, in general, representation with respect to *intersections* of two or more features guarantees higher social welfare than that achieved by the status quo of proportionally representing individual features. Moreover, an analysis of real data suggests that representation with respect to pairs of features is feasible in practice. These results have significant implications for the design of *citizens' assemblies*, which are gaining prominence in AI governance.

Large Vision-Language Models (LVLMs) have demonstrated impressive capabilities in tasks requiring multimodal understanding. However, recent studies indicate that LVLMs are more vulnerable than LLMs to unsafe inputs and prone to generating harmful content. Existing defense strategies primarily include fine-tuning, input sanitization, and output intervention. Although these approaches provide a certain level of protection, they tend to be resource-intensive and struggle to effectively counter sophisticated attack techniques. To tackle such issues, we propose One-head Defense (Oh Defense), a novel yet simple approach utilizing LVLMs' internal safety capabilities. Through systematic analysis of the attention mechanisms, we discover that LVLMs' safety capabilities are concentrated within specific attention heads that respond differently to safe or unsafe inputs. Further exploration reveals that a single critical attention head can effectively serve as a safety guard, providing a strong discriminative signal that amplifies the model's inherent safety capabilities. Hence, the Oh Defense requires no additional training or external modules, making it computationally efficient while effectively reactivating suppressed safety mechanisms. Extensive experiments across diverse LVLM architectures and unsafe datasets validate our approach, i.e., the Oh Defense achieves near-perfect defense success rates (> 98\%) for unsafe inputs while maintaining low false positive rates (< 5\%) for safe content. The source code is available at https://github.com/AIASLab/Oh-Defense.

We examine hypothesis testing within a principal-agent framework, where a strategic agent, holding private beliefs about the effectiveness of a product, submits data to a principal who decides on approval. The principal employs a hypothesis testing rule, aiming to pick a p-value threshold that balances false positives and false negatives while anticipating the agent’s incentive to maximize expected profitability. Building on prior work, we develop a game-theoretic model that captures how the agent’s participation and reporting behavior respond to the principal’s statistical decision rule. Despite the complexity of the interaction, we show that the principal's errors exhibit clear monotonic behavior when segmented by an efficiently computable critical p-value threshold, leading to an interpretable characterization of their optimal p-value threshold. We empirically validate our model and these insights using publicly available data on drug approvals. Overall, our work offers a comprehensive perspective on strategic interactions within the hypothesis testing framework, providing technical and regulatory insights.

The rapid growth of publicly available data has fueled deep learning advancements but also raises concerns about unauthorized data usage. Unlearnable Examples (UEs) have emerged as a data protection strategy that introduces imperceptible perturbations to prevent unauthorized learning. However, most existing UE methods produce perturbations strongly tied to specific training sets, leading to a significant drop in unlearnability when applied to unseen data or tasks. In this paper, we argue that for broad applicability, UEs should maintain their effectiveness across diverse application scenarios. To this end, we conduct the first comprehensive study on the transferability of UEs across diverse and practical yet demanding settings. Specifically, we identify key scenarios that pose significant challenges for existing UE methods, including varying styles, out-of-distribution classes, resolutions, and architectures. Moreover, we propose $\textbf{Versatile Transferable Generator}$ (VTG), a transferable generator designed to safeguard data across various conditions. Specifically, VTG integrates Adversarial Domain Augmentation (ADA) into the generator’s training process to synthesize out-of-distribution samples, thereby improving its generalizability to unseen scenarios. Furthermore, we propose a Perturbation-Label Coupling (PLC) mechanism that leverages contrastive learning to directly align perturbations with class labels. This approach reduces the generator’s reliance on data semantics, allowing VTG to produce unlearnable perturbations in a distribution-agnostic manner. Extensive experiments demonstrate the effectiveness and broad applicability of our approach. Code is available at https://github.com/zhli-cs/VTG.

Stability guarantees have emerged as a principled way to evaluate feature attributions, but existing certification methods rely on heavily smoothed classifiers and often produce conservative guarantees. To address these limitations, we introduce soft stability and propose a simple, model-agnostic, sample-efficient stability certification algorithm (SCA) that yields non-trivial and interpretable guarantees for any attribution method. Moreover, we show that mild smoothing achieves a more favorable trade-off between accuracy and stability, avoiding the aggressive compromises made in prior certification methods. To explain this behavior, we use Boolean function analysis to derive a novel characterization of stability under smoothing. We evaluate SCA on vision and language tasks and demonstrate the effectiveness of soft stability in measuring the robustness of explanation methods.

Deepfake detection models face two critical challenges: generalization to unseen manipulations and demographic fairness among population groups. However, existing approaches often demonstrate that these two objectives are inherently conflicting, revealing a trade-off between them. In this paper, we, for the first time, uncover and formally define a causal relationship between fairness and generalization. Building on the back-door adjustment, we show that controlling for confounders (data distribution and model capacity) enables improved generalization via fairness interventions. Motivated by this insight, we propose Demographic Attribute-insensitive Intervention Detection (DAID), a plug-and-play framework composed of: i) Demographic-aware data rebalancing, which employs inverse-propensity weighting and subgroup-wise feature normalization to neutralize distributional biases; and ii) Demographic-agnostic feature aggregation, which uses a novel alignment loss to suppress sensitive-attribute signals. Across three cross-domain benchmarks, DAID consistently achieves superior performance in both fairness and generalization compared to several state-of-the-art detectors, validating both its theoretical foundation and practical effectiveness.

Multimodal Large Language Models (MLLMs) are increasingly deployed in fine-tuning-as-a-service (FTaaS) settings, where user-submitted datasets adapt general-purpose models to downstream tasks. This flexibility, however, introduces serious security risks, as malicious fine-tuning can implant backdoors into MLLMs with minimal effort. In this paper, we observe that backdoor triggers systematically disrupt cross-modal processing by causing abnormal attention concentration on non-semantic regions—a phenomenon we term **attention collapse**. Based on this insight, we propose **Believe Your Eyes (BYE)**, a data filtering framework that leverages attention entropy patterns as self-supervised signals to identify and filter backdoor samples. BYE operates via a three-stage pipeline: (1) extracting attention maps using the fine-tuned model, (2) computing entropy scores and profiling sensitive layers via bimodal separation, and (3) performing unsupervised clustering to remove suspicious samples. Unlike prior defenses, BYE equires no clean supervision, auxiliary labels, or model modifications. Extensive experiments across various datasets, models, and diverse trigger types validate BYE's effectiveness: it achieves near-zero attack success rates while maintaining clean-task performance, offering a robust and generalizable solution against backdoor threats in MLLMs.

As AI models are deployed with increasing autonomy, it is important to ensure they do not take harmful actions unnoticed. As a potential mitigation, we investigate Chain-of-Thought (CoT) monitoring, wherein a weaker trusted monitor model continuously oversees the intermediate reasoning steps of a more powerful but untrusted model. We compare CoT monitoring to action-only monitoring, where only final outputs are reviewed, in a red-teaming setup where the untrusted model is instructed to pursue harmful side tasks while completing a coding problem. We find that while CoT monitoring is more effective than overseeing only model outputs in scenarios where action-only monitoring fails to reliably identify sabotage, reasoning traces can contain misleading rationalizations that deceive the CoT monitors, reducing performance in obvious sabotage cases. To address this, we introduce a hybrid protocol that independently scores model reasoning and actions, and combines them using a weighted average. Our hybrid monitor consistently outperforms both CoT and action-only monitors across all tested models and tasks, with detection rates twice higher than action-only monitoring for subtle deception scenarios.

We propose a method for discovering and visualizing the differences between two learned representations, enabling more direct and interpretable model comparisons. We validate our method, which we call Representational Differences Explanations (RDX), by using it to compare models with known conceptual differences and demonstrate that it recovers meaningful distinctions where existing explainable AI (XAI) techniques fail. Applied to state-of-the-art models on challenging subsets of the ImageNet and iNaturalist datasets, RDX reveals both insightful representational differences and subtle patterns in the data. Although comparison is a cornerstone of scientific analysis, current tools in machine learning, namely post hoc XAI methods, struggle to support model comparison effectively. Our work addresses this gap by introducing an effective and explainable tool for contrasting model representations.

Despite the remarkable capabilities of large language models (LLMs) across diverse applications, they remain vulnerable to generating content that violates safety regulations and policies. To mitigate these risks, LLMs undergo safety alignment; however, they can still be effectively jailbroken. Off-the-shelf guardrail models are commonly deployed to monitor generations, but these models primarily focus on detection rather than ensuring safe decoding of LLM outputs. Moreover, existing efforts lack rigorous safety guarantees, which are crucial for the universal deployment of LLMs and certifiable compliance with regulatory standards. In this paper, we propose a Claim-based Stream Decoding (CSD) algorithm coupled with a statistical risk guarantee framework using conformal analysis. Specifically, our CSD algorithm integrates a stream guardrail model to safeguard sequential claims generated by LLMs and incorporates a backtracking mechanism to revise claims flagged with high safety risks. We provide theoretical guarantees demonstrating that the CSD algorithm achieves the desired generation distribution subject to safety constraints. Furthermore, we introduce a generation risk certification framework and derive a high-probability upper bound on the safety risk of the proposed CSD algorithm. We extend our approach to online settings, where user queries arrive sequentially, and prove that our method can asymptotically control safety risk to any desired level. Empirical evaluations demonstrate the effectiveness and efficiency of the CSD algorithm compared to state-of-the-art safety decoding approaches. Additionally, we validate the soundness and tightness of the derived safety risk upper bound using realistic data in both offline and online scenarios.

Influence Maximization (IM) seeks a seed set to maximize information dissemination in a network. Elegant IM algorithms could naturally extend to cases where each node is equipped with a specific weight, reflecting individual gains to measure its importance. In prevailing literature, these gains are typically assumed to remain constant throughout diffusion and are solvable through explicit formulas based on node characteristics and network topology. However, this assumption is not always feasible due to two key challenges: 1) \textit{Unobservability}: The individual gains of each node are primarily evaluated by the difference between the outputs in the activated and non-activated states. In practice, we can only observe one of these states, with the other remaining unobservable post-propagation. 2) \textit{Environmental sensitivity}: Beyond nodes’ inherent properties, individual gains are also sensitive to the activation status of surrounding nodes, which change dynamically during propagation even when the network topology is fixed. To address these uncertainties, we introduce a Causal Influence Maximization (CauIM) framework, leveraging causal inference techniques to model dynamic individual gains. We propose two algorithms, G-CauIM and A-CauIM, where the latter incorporates a novel acceleration technique. Theoretically, we establish the generalized lower bound of influence spread and provide robustness analysis. Empirically, experiments on synthetic and real-world datasets validate the effectiveness and reliability of our approach.