Who Bridges Safety? Identifying and Targeting Cross-Lingual Shared Safety Pathways

Uncovering the internal mechanisms underlying the safety capabilities of large language models (LLMs) is crucial for developing trustworthy artificial intelligence. Currently, mechanistic interpretability studies on multilingual safety are largely confined to local components, such as isolated neurons. However, this static and fragmented perspective overlooks the synergy among components and fails to elucidate how safety signals dynamically propagate within the model to drive safety decisions ultimately. In this work, we move beyond isolated neurons to identify and target the cross-layer functional pathways formed during safety signal propagation, thereby uncovering the mechanisms driving the cross-lingual safety gap. Specifically, we first identify monolingual safety pathways and validate their impact on refusing harmful requests. Subsequent cross-lingual analyses reveal a sparse subset of cross-lingual shared safety pathways, confirming that this intersection acts as the internal bridge transferring safety capabilities from high-resource (HR) languages to non-high-resource (NHR) languages. Building on these mechanistic findings, we propose a pathways-targeted alignment method based on the cross-lingual shared safety pathways. Experimental results show that updating only a small fraction of pathway parameters significantly improves safety in NHR languages while largely preserving the model's general capabilities.

Towards Privacy-Preserving Federated Prompt Tuning under Data Heterogeneity: A Subspace-Decomposed Expert Approach

Federated prompt tuning (FPT) enables collaborative adaptation of vision--language models (VLMs) using lightweight prompts. Existing methods often address heterogeneity and privacy through a split-prompt design under local differential privacy (DP), combining a shared prompt for global transfer with private prompts for local adaptation. However, a single shared prompt may over-smooth diverse transferable knowledge, weakening the balance between personalization and generalization. Multi-expert prompts (MEPs) can better capture this diversity, but enlarge the communicated space, increasing DP noise and communication cost while making robust expert composition more difficult. We propose FedSEPT, a privacy-preserving Fed}erated Subspace-decomposed Expert Prompt Tuning. Specifically, we employ Subspace-decomposed Expert Modeling (SEM) to parameterize multiple prompt experts with shared low-rank factors, a fixed public basis, and private residuals, thereby confining communication and DP perturbation to a compact factor space while enabling direct server aggregation in a common coordinate system. We further design Instance-aware Expert Fusion (IEF), which adaptively combines semantically complementary experts via on-device routing and performs efficient logit-level fusion using cached expert-specific text features. Extensive experiments on 11 heterogeneous benchmarks show that, under the same privacy constraints, FedSEPT achieves a better trade-off between local adaptation and global generalization than strong baselines.

AURORA: Asymmetry and Update-Induced Rotation for Robust Hallucination Detection in Large Language Models

Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide range of natural language processing tasks. However, their tendency to generate hallucinations, namely factually incorrect or unfaithful outputs, poses a critical obstacle to their deployment in high-stakes applications. Although recent hallucination detection methods have made encouraging progress, they typically rely on costly output-level consistency checks or static hidden-state probes that capture shallow dataset-specific patterns, leading to substantial degradation under cross-dataset evaluation. In this work, we propose AURORA, a novel hallucination detection framework that shifts the focus from static representations to the weight-gradient dynamics of LLMs. Our key insight is that hallucinated and faithful answers induce qualitatively different gradient update patterns on the model's parameters. Specifically, hallucinated samples trigger asymmetric and structurally misaligned gradients, which can be captured through two complementary features: (1) the skewness of the cosine similarity distribution between weight matrices and their gradient update directions, and (2) the rotation ratio, which quantifies how much the gradient update reorients the singular-vector basis of weight matrices via SVD. AURORA achieves strong hallucination detection performance across four model families and four benchmark datasets. Further analyses demonstrate that our method scales effectively across model sizes and transfers to out-of-domain tasks, including mathematical reasoning and vision-language scenarios.

Mosaic: Multimodal Jailbreak against Closed-Source VLMs via Multi-View Ensemble Optimization

Vision-Language Models (VLMs) are powerful but remain vulnerable to multimodal jailbreak attacks. Existing attacks mainly rely on either explicit visual prompt attacks or gradient-based adversarial optimization. While the former is easier to detect, the latter produces subtle perturbations that are less perceptible, but is usually optimized and evaluated under homogeneous open-source surrogate-target settings, leaving its effectiveness on commercial closed-source VLMs under heterogeneous settings unclear. To examine this issue, we study different surrogate-target settings and observe a consistent gap between homogeneous and heterogeneous settings, a phenomenon we term surrogate dependency. Motivated by this finding, we propose Mosaic, a Multi-view ensemble optimization framework for multimodal jailbreak against closed-source VLMs, which alleviates surrogate dependency under heterogeneous surrogate-target settings by reducing over-reliance on any single surrogate model and visual view. Specifically, Mosaic incorporates three core components: a Text-Side Transformation module, which perturbs refusal-sensitive lexical patterns; a Multi-View Image Optimization module, which updates perturbations under diverse cropped views to avoid overfitting to a single visual view; and a Surrogate Ensemble Guidance module, which aggregates optimization signals from multiple surrogate VLMs to reduce surrogate-specific bias. Extensive experiments on safety benchmarks demonstrate that Mosaic achieves state-of-the-art Attack Success Rate and Average Toxicity against commercial closed-source VLMs.

From Unfamiliar to Familiar: Detecting Pre-training Data via Gradient Deviations in Large Language Models

Pre-training data detection for LLMs is essential for addressing copyright concerns and mitigating benchmark contamination. Existing methods mainly focus on the likelihood-based statistical features or heuristic signals before and after fine-tuning, but the former are susceptible to word frequency bias in corpora, and the latter strongly depend on the similarity of fine-tuning data. From an optimization perspective, we observe that during training, samples transition from unfamiliar to familiar in a manner reflected by systematic differences in gradient behavior. Familiar samples exhibit smaller update magnitudes, distinct update locations in model components, and more sharply activated neurons. Based on this insight, we propose GDS, a method that identifies pre-training data by probing Gradient Deviation Scores of target samples. Specifically, we first represent each sample using gradient profiles that capture the magnitude, location, and concentration of parameter updates across FFN and Attention modules, revealing consistent distinctions between member and non-member data. These features are then fed into a lightweight classifier to perform binary membership inference. Experiments on five public datasets show that GDS achieves state-of-the-art performance with significantly improved cross-dataset transferability over strong baselines. Further interpretability analyse show gradient feature distribution differences, enabling practical and scalable pre-training data detection.

Replacing Parameters with Preferences: Federated Alignment of Heterogeneous Vision-Language Models

VLMs have broad potential in privacy-sensitive domains such as healthcare and finance, yet strict data-sharing constraints render centralized training infeasible. FL mitigates this issue by enabling decentralized training, but practical deployments face challenges due to client heterogeneity in computational resources, application requirements, and model architectures. We argue that while replacing data with model parameters characterizes the present of FL, replacing parameters with preferences represents a more scalable and privacy-preserving future. Motivated by this perspective, we propose MoR, a federated alignment framework based on GRPO with Mixture-of-Rewards for heterogeneous VLMs. MoR initializes a visual foundation model as a KL-regularized reference, while each client locally trains a reward model from local preference annotations, capturing specific evaluation signals without exposing raw data. To reconcile heterogeneous rewards, we introduce a routing-based fusion mechanism that adaptively aggregates client reward signals. Finally, the server performs GRPO with this mixed reward to optimize the base VLM. Experiments on three public VQA benchmarks demonstrate that MoR consistently outperforms federated alignment baselines in generalization, robustness, and cross-client adaptability. Our approach provides a scalable solution for privacy-preserving alignment of heterogeneous VLMs under federated settings.

Stable-RAG: Mitigating Retrieval-Permutation-Induced Hallucinations in Retrieval-Augmented Generation

Retrieval-Augmented Generation (RAG) has become a key paradigm for reducing factual hallucinations in large language models (LLMs), yet little is known about how the order of retrieved documents affects model behavior. We empirically show that under Top-5 retrieval with the gold document included, LLM answers vary substantially across permutations of the retrieved set, even when the gold document is fixed in the first position. This reveals a previously underexplored sensitivity to retrieval permutations. Although robust RAG methods primarily focus on enhancing LLM robustness to low-quality retrieval and mitigating positional bias to distribute attention fairly over long contexts, neither approach directly addresses permutation sensitivity. In this paper, we propose Stable-RAG, which exploits permutation sensitivity estimation to mitigate permutation-induced hallucinations. Stable-RAG runs the generator under multiple retrieval orders, clusters hidden states, and decodes from a cluster-center representation that captures the dominant reasoning pattern. It then uses these reasoning results to align hallucinated outputs toward the correct answer, encouraging the model to produce consistent and accurate predictions across document permutations. Experiments on three QA datasets show that Stable-RAG significantly improves answer accuracy, reasoning consistency and robust generalization across datasets, retrievers, and input lengths compared with baselines.

Parameter Importance-Driven Continual Learning for Foundation Models

Domain-specific post-training often causes catastrophic forgetting, making foundation models lose their general reasoning ability and limiting their adaptability to dynamic real-world environments. Preserving general capabilities while acquiring downstream domain knowledge is a central challenge for large language and multimodal models. Traditional continual learning methods, such as regularization, replay and architectural isolation, suffer from poor downstream performance, reliance on inaccessible historical data, or additional parameter overhead. While recent parameter-efficient tuning (PET) methods can alleviate forgetting, their effectiveness strongly depends on the choice of parameters and update strategies. In this paper, we introduce PIECE, a Parameter Importance Estimation-based Continual Enhancement method that preserves general ability while efficiently learning domain knowledge without accessing prior training data or increasing model parameters. PIECE selectively updates only 0.1% of core parameters most relevant to new tasks, guided by two importance estimators: PIECE-F based on Fisher Information, and PIECE-S based on a second-order normalization that combines gradient and curvature information. Experiments across three language models and two multimodal models show that PIECE maintains general capabilities and achieves state-of-the-art continual learning performance across diverse downstream tasks. Our results highlight a practical path to scalable, domain-adaptive foundation models without catastrophic forgetting.

Unlocking Dynamic Inter-Client Spatial Dependencies: A Federated Spatio-Temporal Graph Learning Method for Traffic Flow Forecasting

Spatio-temporal graphs are powerful tools for modeling complex dependencies in traffic time series. However, the distributed nature of real-world traffic data across multiple stakeholders poses significant challenges in modeling and reconstructing inter-client spatial dependencies while adhering to data locality constraints. Existing methods primarily address static dependencies, overlooking their dynamic nature and resulting in suboptimal performance. In response, we propose Federated Spatio-Temporal Graph with Dynamic Inter-Client Dependencies (FedSTGD), a framework designed to model and reconstruct dynamic inter-client spatial dependencies in federated learning. FedSTGD incorporates a federated nonlinear computation decomposition module to approximate complex graph operations. This is complemented by a graph node embedding augmentation module, which alleviates performance degradation arising from the decomposition. These modules are coordinated through a client-server collective learning protocol, which decomposes dynamic inter-client spatial dependency learning tasks into lightweight, parallelizable subtasks. Extensive experiments on four real-world datasets demonstrate that FedSTGD achieves superior performance over state-of-the-art baselines in terms of RMSE, MAE, and MAPE, approaching that of centralized baselines. Ablation studies confirm the contribution of each module in addressing dynamic inter-client spatial dependencies, while sensitivity analysis highlights the robustness of FedSTGD to variations in hyperparameters.

Detecting Stealthy Backdoor Samples based on Intra-class Distance for Large Language Models

Fine-tuning LLMs with datasets containing stealthy backdoors from publishers poses security risks to downstream applications. Mainstream detection methods either identify poisoned samples by analyzing the prediction probability of poisoned classification models or rely on the rewriting model to eliminate the stealthy triggers. However, the former cannot be applied to generation tasks, while the latter may degrade generation performance and introduce new triggers. Therefore, efficiently eliminating stealthy poisoned samples for LLMs remains an urgent problem. We observe that after applying TF-IDF clustering to the sample response, there are notable differences in the intra-class distances between clean and poisoned samples. Poisoned samples tend to cluster closely because of their specific malicious outputs, whereas clean samples are more scattered due to their more varied responses. Thus, in this paper, we propose a stealthy backdoor sample detection method based on Reference-Filtration and Tfidf-Clustering mechanisms (RFTC). Specifically, we first compare the sample response with the reference model's outputs and consider the sample suspicious if there's a significant discrepancy. And then we perform TF-IDF clustering on these suspicious samples to identify the true poisoned samples based on the intra-class distance. Experiments on two machine translation datasets and one QA dataset demonstrate that RFTC outperforms baselines in backdoor detection and model performance. Further analysis of different reference models also confirms the effectiveness of our Reference-Filtration.