Large language model (LLM) alignment faces a critical dilemma when addressing multiple human preferences: improvements in one dimension frequently come at the expense of others, creating unavoidable trade-offs between competing objectives like helpfulness and harmlessness. While prior work mainly focuses on constraint-based optimization algorithms and data selection strategies to mitigate conflicts, these approaches overlook the fundamental issue of resolving conflicts directly at the parameter level. In this paper, we present OrthAlign, an innovative approach that pioneers a new paradigm by leveraging orthogonal subspace decomposition to fundamentally resolve gradient-level conflicts in multi-objective preference alignment. OrthAlign strategically decomposes parameter update spaces into orthogonal subspaces, ensuring that optimization toward different preferences occurs in mathematically non-interfering directions. Building upon this, we provide theoretical guarantees demonstrating that when parameter increments satisfy both orthogonal subspace constraints and spectral norm bounds, the resulting updates exhibit linear Lipschitz growth rather than exponential instability, ensuring stable convergence across all preference dimensions. Extensive experiments show that: I. OrthAlign achieves maximum single-preference improvements ranging from 34.61% to 50.89% after multiple-objective alignment across helpful, harmless, and truthful dimensions. II. With an average overall reward improvement of 13.96%. Our code is available at https://anonymous.4open.science/r/OrthAlign.

Large Language Model-based Multi-Agent Systems (MASs) have demonstrated strong advantages in addressing complex real-world tasks. However, due to the introduction of additional attack surfaces, MASs are particularly vulnerable to misinformation injection. To facilitate a deeper understanding of misinformation propagation dynamics within these systems, we introduce **MisinfoTask**, a novel dataset featuring complex, realistic tasks designed to evaluate MAS robustness against such threats. Building upon this, we propose **ARGUS**, a two-stage, training-free defense framework leveraging goal-aware reasoning for precise misinformation rectification within information flows. Our experiments demonstrate that in challenging misinformation scenarios, ARGUS exhibits significant efficacy across various injection attacks, achieving an average reduction in misinformation toxicity of approximately 28.17% and improving task success rates under attack by approximately 10.33%.

Spatio-temporal (ST) prediction is crucial in earth sciences, including meteorological forecasting and urban computing, to name just a few. Access to ample high-quality data, combined with deep models adept at inference, is essential for attaining significant outcomes. Yet, data scarcity and the substantial costs of sensor deployment result in notable data imbalances. Overly specialized models that lack causal linkages further undermine the generalizability of inference techniques. To address these challenges, we first introduce a causal framework for ST predictions, named $\mathtt{NuwaDynamics}$NuwaDynamicsNuwaDynamics, aimed at pinpointing causal regions in data and providing models with the capability for causal reasoning in a dual-phase process. Initially, we employ upstream self-supervision to identify causally significant patches, equipping the model with generalizable insights and performing targeted interventions on non-essential patches to approximate potential testing distributions. This stage is known as the discovery phase. Progressing from discovery, we apply the insights to downstream tasks tailored to specific ST goals, enhancing the model's recognition of a wider potential data distribution and augmenting its causal perceptual abilities (referred to as the Update phase). Additionally, we address environmental controllability and high computational complexity by implementing channel multiplication and conditional generation methods. This process, termed $\mathtt{NuwaDynamics+}$NuwaDynamics+NuwaDynamics+, can further be interpreted as the front-door adjustment technique in the causality domain. Through comprehensive experiments across ten real-world or simulated ST benchmarks, we demonstrate that integrating the $\mathtt{NuwaDynamics+}$NuwaDynamics+NuwaDynamics+ concept substantially improves various model performance. $\mathtt{NuwaDynamics+}$NuwaDynamics+NuwaDynamics+ concept also significantly enhances the versatility across various dynamic ST tasks, such as extreme weather forecasting and long-temporal-step super-resolution predictions.

Embodied Artificial Intelligence (Embodied AI) integrates perception, cognition, planning, and interaction into agents that operate in open-world, safety-critical environments. As these systems gain autonomy and enter domains such as transportation, healthcare, and industrial or assistive robotics, ensuring their safety becomes both technically challenging and socially indispensable. Unlike digital AI systems, embodied agents must act under uncertain sensing, incomplete knowledge, and dynamic human-robot interactions, where failures can directly lead to physical harm. This survey provides a comprehensive and structured review of safety research in embodied AI, examining attacks and defenses across the full embodied pipeline, from perception and cognition to planning, action and interaction, and agentic system. We introduce a multi-level taxonomy that unifies fragmented lines of work and connects embodied-specific safety findings with broader advances in vision, language, and multimodal foundation models. Our review synthesizes insights from over 500 papers spanning adversarial, backdoor, jailbreak, and hardware-level attacks; attack detection, safe training and robust inference; and risk-aware human-agent interaction. This analysis reveals several overlooked challenges, including the fragility of multimodal perception fusion, the instability of planning under jailbreak attacks, and the trustworthiness of human-agent interaction in open-ended scenarios. By organizing the field into a coherent framework and identifying critical research gaps, this survey provides a roadmap for building embodied agents that are not only capable and autonomous but also safe, robust, and reliable in real-world deployment.

The rapid development and widespread adoption of Audio Large Language Models (ALLMs) require a rigorous assessment of their trustworthiness. However, existing evaluation frameworks, primarily designed for text, are not equipped to handle the unique vulnerabilities introduced by audio’s acoustic properties. We find that significant trustworthiness risks in ALLMs arise from non-semantic acoustic cues, such as timbre, accent, and background noise, which can be used to manipulate model behavior. To address this gap, we propose AudioTrust, the first framework for large-scale and systematic evaluation of ALLM trustworthiness concerning these audio-specific risks. AudioTrust spans six key dimensions: fairness, hallucination, safety, privacy, robustness, and authenticition. It is implemented through 26 distinct sub-tasks and a curated dataset of over 4,420 audio samples collected from real-world scenarios (e.g., daily conversations, emergency calls, and voice assistant interactions), purposefully constructed to probe the trustworthiness of ALLMs across multiple dimensions. Our comprehensive evaluation includes 18 distinct experimental configurations and employs human-validated automated pipelines to objectively and scalably quantify model outputs. Experimental results reveal the boundaries and limitations of 14 state-of-the-art (SOTA) open-source and closed-source ALLMs when confronted with diverse high-risk audio scenarios, thereby offering critical insights into the secure and trustworthy deployment of future audio models. Our platform and benchmark are publicly available at https://anonymous.4open.science/r/AudioTrust-8715/.

Seeing clearly with high resolution is a foundation of Multimodal Large Language Models (MLLMs), which has been proven to be vital for visual perception and reasoning. Existing works usually employ a straightforward resolution upscaling method, where the image consists of global and local branches, with the latter being the sliced image patches but resized to the same resolution as the former. This means that higher resolution requires more local patches, resulting in exorbitant computational expenses, and meanwhile, the dominance of local image tokens may diminish the global context. In this paper, we dive into the problems and propose a new framework as well as an elaborate optimization strategy. Specifically, we extract contextual information from the global view using a mixture of adapters, based on the observation that different adapters excel at different tasks. With regard to local patches, learnable query embeddings are introduced to reduce image tokens, the important tokens most relevant to the user question will be further selected by a similarity-based selector. Our empirical results demonstrate a ‘less is more’ pattern, where utilizing fewer but more informative local image tokens leads to improved performance. Besides, a significant challenge lies in the training strategy, as simultaneous end-to-end training of the global mining block and local compression block does not yield optimal results. We thus advocate for an alternating training way, ensuring balanced learning between global and local aspects. Finally, we also introduce a challenging dataset with high requirements for image detail, enhancing the training of the local compression layer. The proposed method, termed MLLM with Sophisticated Tasks, Local image compression, and Mixture of global Experts (SliME), achieves leading performance across various benchmarks with only 2 million training data.

As the field of Multimodal Large Language Models (MLLMs) continues to evolve, their potential to revolutionize artificial intelligence is particularly promising, especially in addressing mathematical reasoning tasks. Current mathematical benchmarks predominantly focus on evaluating MLLMs' problem-solving ability, yet there is a crucial gap in addressing more complex scenarios such as error detection, for enhancing reasoning capability in complicated settings. To fill this gap, we formally formulate the new task: multimodal error detection, and introduce ErrorRadar, the first benchmark designed to assess MLLMs' capabilities in such a task. ErrorRadar evaluates two sub-tasks: error step identification and error categorization, providing a comprehensive framework for evaluating MLLMs' complex mathematical reasoning ability. It consists of 2,500 high-quality multimodal K-12 mathematical problems, collected from real-world student interactions in an educational organization, with rigorous annotation and rich metadata such as problem type and error category. Through extensive experiments, we evaluated both open-source and closed-source representative MLLMs, benchmarking their performance against educational expert evaluators. Results indicate significant challenges still remain, as GPT-4o with best performance is still around 10

Autonomous agents based on large language models (LLMs) have demonstrated impressive capabilities in numerous real-world applications. While most LLMs are limited in several key agentic procedures, graphs can serve as a powerful auxiliary structure to enhance structure, continuity, and coordination in complex agent workflows. Given the rapid growth and fragmentation of research on Graph-augmented LLM Agents (GLA), this article offers a timely and comprehensive overview of recent advances and highlights key directions for future work. Specifically, we categorize existing GLA methods by their primary functions in LLM agent systems, including planning, memory, and tool usage, and then analyze how graphs and graph learning algorithms contribute to each. For multiagent systems, we further discuss how GLA solutions facilitate the orchestration, efficiency optimization, and trustworthiness of MAS. Finally, we highlight key future directions to advance this field, from improving structural adaptability to enabling unified, scalable, and multimodal GLA systems.

Despite the rapid evolution of training paradigms, the decoder backbone of large vision--language models (LVLMs) remains fundamentally rooted in the residual-connection Transformer architecture. Therefore, deciphering the distinct roles of internal modules is critical for understanding model mechanics and guiding architectural optimization. While prior statistical approaches have provided valuable attribution-based insights, they often lack a unified theoretical basis. To bridge this gap, we propose a unified framework grounded in to quantify the of residual updates. Applying this unified framework reveals a fundamental functional decoupling: focused on reconfiguration, whereas driving semantic innovation. Strikingly, further experiments demonstrate that replacing learned attention weights with predefined values (e.g., Gaussian noise) yields comparable or even superior performance across a majority of datasets relative to vanilla models. These results expose severe in current mechanisms, suggesting that state-of-the-art LVLMs effectively ``get lost in attention'' rather than efficiently leveraging visual context.

The rapid advancement of generative AI highlights the importance of text-to-image (T2I) security, particularly with the threat of backdoor poisoning. Timely disclosure and mitigation of security vulnerabilities in T2I models are crucial for ensuring the safe deployment of generative models. We explore a novel training-free backdoor poisoning paradigm through model editing, which is recently employed for knowledge updating in large language models. Nevertheless, we reveal the potential security risks posed by model editing techniques to image generation models. In this work, we establish the principles for backdoor attacks based on model editing, and propose a relationship-driven precise backdoor poisoning method, REDEditing. Drawing on the principles of equivalent-attribute alignment and stealthy poisoning, we develop an equivalent relationship retrieval and joint-attribute transfer approach that ensures consistent backdoor image generation through concept rebinding. A knowledge isolation constraint is proposed to preserve benign generation integrity. Our method achieves an 11% higher attack success rate compared to state-of-the-art approaches. Remarkably, adding just one line of code enhances output naturalness while improving backdoor stealthiness by 24%. This work aims to heighten awareness regarding this security vulnerability in editable image generation models.

Large Language Model (LLM)-based agentic systems, often comprising multiple models, complex tool invocations, and orchestration protocols, substantially outperform monolithic agents. Yet this very sophistication amplifies their fragility, making them more prone to system failure. Pinpointing the specific agent or step responsible for an error within long execution traces defines the task of \textbf{agentic system failure attribution}. Current state-of-the-art reasoning LLMs, however, remain strikingly inadequate for this challenge, with accuracy generally below $10\\%$. To address this gap, we propose AgenTracer, the first automated framework for annotating failed multi-agent trajectories via counterfactual replay and programmed fault injection, producing the curated dataset TracerTraj. Leveraging this resource, we develop AgenTracer-8B, a lightweight failure tracer trained with multi-granular reinforcement learning, capable of efficiently diagnosing errors in verbose multi-agent interactions. On {Who\&When} benchmark, AgenTracer-8B outperforms giant proprietary LLMs like Gemini-2.5-Pro and Claude-4-Sonnet by up $18.18\\%$, setting a new standard in LLM agentic failure attribution. More importantly, AgenTracer-8B delivers actionable feedback to off-the-shelf multi-agent systems like MetaGPT and MaAS with $4.8\sim14.2\\%$ performance gains, empowering self-correcting and self-evolving agentic AI.

GUI agents are beginning to operate the web, mobile, and desktop as interactive worlds, where successful control depends on carrying forward visual, procedural, and task-level evidence beyond the fleeting present screen. Yet most agents still treat memory as an external, human-readable artifact: histories are summarized, categorized, retrieved, and reinserted as text or structured records before being encoded again by the policy. This creates a mismatch between the representational form in which experience is stored and the latent embedding sequence over which modern GUI policies actually act. We introduce Mem-W, a series of latent-memory-native GUI agents that treat memory as part of the agent's continuous context rather than as an auxiliary symbolic scaffold. Mem-W weaves both historical trajectories (as experiential memory) and in-session segments (as working memory) into compact memory tokens through a shared trajectory-to-latent compressor. These tokens are woven with the current GUI observation and local context into one continuous embedding sequence, allowing the agent to read successes, failures, and unfinished progress through the same machine-native interface. Mem-W is trained with self-distillation and outcome-aware supervision to preserve decision-relevant state while filtering memory toward evidence that truly supports task success. Across four web and mobile navigation benchmarks, Mem-W consistently improves diverse backbones and memory-enhanced baselines, with gains of up to +30.0, suggesting that latent-context-native memory can serve as a scalable foundation for long-horizon GUI agency.

Missed and delayed diagnosis remains a major challenge in rare disease care. At the initial clinical encounters, physicians assess rare disease risk using only limited information under high uncertainty. When high-risk patients are not recognised at this stage, targeted diagnostic testing is often not initiated, resulting in missed diagnosis. Existing primary care triage processes are structurally insufficient to reliably identify patients with rare diseases at initial clinical presentation and universal screening is needed to reduce diagnostic delay. Here we present RareAlert, an early screening system which predict patient-level rare disease risk from routinely available primary-visit information. RareAlert integrates reasoning generated by ten LLMs, calibrates and weights these signals using machine learning, and distils the aligned reasoning into a single locally deployable model. To develop and evaluate RareAlert, we curated RareBench, a real-world dataset of 158,666 cases covering 33 Orphanet disease categories and more than 7,000 rare conditions, including both rare and non-rare presentations. The results showed that rare disease identification can be reconceptualised as a universal uncertainty resolution process applied to the general patient population. On an independent test set, RareAlert, a Qwen3-4B based model trained with calibrated reasoning signals, achieved an AUC of 0.917, outperforming the best machine learning ensemble and all evaluated LLMs, including GPT-5, DeepSeek-R1, Claude-3.7-Sonnet, o3-mini, Gemini-2.5-Pro, and Qwen3-235B. These findings demonstrate the diversity in LLM medical reasoning and the effectiveness of aligning such reasoning in highly uncertain clinical tasks. By incorporating calibrated reasoning into a single model, RareAlert enables accurate, privacy-preserving, and scalable rare disease risk screening suitable for large-scale local deployment.

Explainability in Graph Neural Networks (GNNs) has shown considerable promise in bolstering their trustworthiness, credibility and transparency. Our research delves into the assessment of explainability within GNNs, a pivotal factor for ensuring the reliability of explainability techniques in real-world applications. Existing evaluation metrics, typically involving taking explanatory subgraphs as inputs and measuring output differences, often face out-of-distribution (OOD) challenges. This issue occurs when explanatory subgraphs do not align with real-world data distributions, affecting the reliability of model explanations. With this in mind, in this work, we endeavor to confront this issue by introducing a novel evaluation metric, termed OOD-resistant Adversarial Robustness (OAR). Specifically, our approach is inspired by adversarial robustness, assessing the resilience of explanation subgraphs to attacks. Additionally, we incorporate a sophisticated OOD reweighting mechanism within the evaluation framework to ensure that assessments remain aligned with the original data distribution. Going beyond this, to accommodate a wider range of evaluation tasks, we further devise a counterfactual attack module and complement the perturbed subgraph using the conditional graph diffusion model. The refined paradigm, termed OAR+, ensures that our metric is versatile and applicable across various contexts. Furthermore, we establish a standardized framework, which serves as a benchmark for evaluating the fairness and accuracy of different metrics. We conduct extensive experiments to validate the effectiveness of the OAR and OAR+.

Latent space is rapidly emerging as a native substrate for language-based models. While modern systems are still commonly understood through explicit token-level generation, an increasing body of work shows that many critical internal processes are more naturally carried out in continuous latent space than in human-readable verbal traces. This shift is driven by the structural limitations of explicit-space computation, including linguistic redundancy, discretization bottlenecks, sequential inefficiency, and semantic loss. This survey aims to provide a unified and up-to-date landscape of latent space in language-based models. We organize the survey into five sequential perspectives: Foundation, Evolution, Mechanism, Ability, and Outlook. We begin by delineating the scope of latent space, distinguishing it from explicit or verbal space and from the latent spaces commonly studied in generative visual models. We then trace the field's evolution from early exploratory efforts to the current large-scale expansion. To organize the technical landscape, we examine existing work through the complementary lenses of mechanism and ability. From the perspective of Mechanism, we identify four major lines of development: Architecture, Representation, Computation, and Optimization. From the perspective of Ability, we show how latent space supports a broad capability spectrum spanning Reasoning, Planning, Modeling, Perception, Memory, Collaboration, and Embodiment. Beyond consolidation, we discuss the key open challenges, and outline promising directions for future research. We hope this survey serves not only as a reference for existing work, but also as a foundation for understanding latent space as a general computational and systems paradigm for next-generation intelligence.

Vision-Language Models (VLMs) frequently suffer from visual perception errors and hallucinations that compromise answer accuracy in complex reasoning tasks. Reinforcement Learning with Verifiable Rewards (RLVR) offers a promising solution by optimizing policies using answer correctness signals. Despite their effectiveness, prevailing RLVR methods face two critical limitations. First, much of the sampling budget is wasted on trajectories doomed to fail due to early visual description errors. Second, sparse rewards cannot distinguish whether failures stem from visual perception or reasoning stages. We introduce MIRL, a decoupled framework that addresses both limitations by leveraging mutual information (MI) between generated descriptions and visual inputs as a cheap pre-screening signal. This enables intelligent budget allocation toward high-potential trajectories via forking, while decoupled training provides independent MI-based rewards for visual perception optimization, resolving reward blindness. Experiments on six vision-language reasoning benchmarks demonstrate that MIRL achieves 70.22

Backdoor attacks are a significant threat to large language models (LLMs), often embedded via public checkpoints, yet existing defenses rely on impractical assumptions about trigger settings. To address this challenge, we propose Locphylax, a defense framework that requires no prior knowledge of trigger settings. Locphylax is based on the key observation that when deliberately injecting known backdoors into an already-compromised model, both existing unknown and newly injected backdoors aggregate in the representation space. Locphylax leverages this through a two-stage process: first, aggregating backdoor representations by injecting known triggers, and then, performing recovery fine-tuning to restore benign outputs. Extensive experiments across multiple LLM architectures demonstrate that: (I) Locphylax reduces the average Attack Success Rate to 4.41% across multiple benchmarks, outperforming existing baselines by 28.1%–69.3%. (II) Clean accuracy and utility are preserved within 0.5% of the original model, ensuring negligible impact on legitimate tasks. (III) The defense generalizes across different types of backdoors, confirming its robustness in practical deployment scenarios. Our code is available at https://anonymous.4open.science/r/Locphylax.

LLM-based multi-agent systems exhibit strong collaborative capabilities but often suffer from redundant communication and excessive token overhead. Existing methods typically enhance efficiency through pretrained GNNs or greedy algorithms, but often isolate pre- and post-task optimization, lacking a unified strategy. To this end, we present SafeSieve, a progressive and adaptive multi-agent pruning algorithm that dynamically refines the inter-agent communication through a novel dual-mechanism. SafeSieve integrates initial LLM-based semantic evaluation with accumulated performance feedback, enabling a smooth transition from heuristic initialization to experience-driven refinement. Unlike existing greedy Top-k pruning methods, SafeSieve employs 0-extension clustering to preserve structurally coherent agent groups while eliminating ineffective links. Experiments across benchmarks (SVAMP, HumanEval, etc.) showcase that SafeSieve achieves 94.01% average accuracy while reducing token usage by 12.4%-27.8%. Results further demonstrate robustness under prompt injection attacks (1.23% average accuracy drop). In heterogeneous settings, SafeSieve reduces deployment costs by 13.3% while maintaining performance. These results establish SafeSieve as an efficient, GPU-free, and scalable framework for practical multi-agent systems. Our code can be found below.

Recent advances in multimodal large language models largely rely on CLIP-based visual encoders, which emphasize global semantic alignment but struggle with fine-grained visual understanding. In contrast, DINOv3 provides strong pixel-level perception yet lacks coarse-grained semantic abstraction, leading to limited multi-granularity reasoning. To address this gap, we propose Granulon, a novel DINOv3-based MLLM with adaptive granularity augmentation. Granulon introduces a text-conditioned granularity Controller that dynamically adjusts the visual abstraction level according to the semantic scope of the textual input, and an Adaptive Token Aggregation module that performs granularity-guided pooling and relation-aware clustering to produce compact, semantically rich visual tokens. This design enables unified "pixel-to-fine-to-coarse" reasoning within a single forward pass. Extensive and interpretable experiments demonstrate that Granulon improves accuracy by 30% and reduces hallucination by 20%, outperforming all visual encoders under identical settings. Code is available at the Supplementary.

Urban socioeconomic indicator prediction aims to infer various metrics related to sustainable development in diverse urban landscapes using data-driven methods. However, prevalent pretrained models, particularly those reliant on satellite imagery, face dual challenges. Firstly, concentrating solely on macro-level patterns from satellite data may introduce bias, lacking nuanced details at micro levels, such as architectural details at a place. Secondly, the text generated by the precursor work UrbanCLIP, which fully utilizes the extensive knowledge of LLMs, frequently exhibits issues such as hallucination and homogenization, resulting in a lack of reliable quality. In response to these issues, we devise a novel framework entitled UrbanVLP based on Vision-Language Pretraining. Our UrbanVLP seamlessly integrates multi-granularity information from both macro (satellite) and micro (street-view) levels, overcoming the limitations of prior pretrained models. Moreover, it introduces automatic text generation and calibration, providing a robust guarantee for producing high-quality text descriptions of urban imagery. Rigorous experiments conducted across six socioeconomic indicator prediction tasks underscore its superior performance.