Yanshu Li

Jiaqian Li, Yanshu Li, Boxuan Zhang et al.

LLM agents increasingly operate through multi-turn tool use and environment interaction, where safety risks often emerge from intermediate steps long before they surface in the final outcome. Reactive auditing is therefore insufficient: post-hoc diagnosis frequently misses the chance to flag risks while they are unfolding. We propose TRACES, a representation-based proactive auditor that learns prefix-level trajectory risk states from the hidden representations of an observer LLM. TRACES induces latent mechanism features from step representations and models their temporal evolution to estimate whether a partial trajectory is drifting toward unsafe behavior. To sidestep the cost and ambiguity of step-level risk annotation, TRACES is trained with weak trajectory-level supervision while still producing dense prefix-level risk estimates. Across multiple agent safety benchmarks, TRACES improves both full-trajectory safety prediction and proactive risk discrimination. Our analyses further suggest that these risk states can help train a safer agent, highlighting the broader potential of proactive auditing for long-horizon agent safety.

Yanshu Li, Jiaqian Li, Kuai Yu et al.

Large vision-language models (LVLMs) have demonstrated strong general multimodal capability and are increasingly deployed in downstream systems. This trend has driven growing interest in LVLM personalization, which aims to enable models to quickly and effectively learn out-of-distribution multimodal concepts to meet user-specific needs. However, many existing methods rely on inference-time training, which reduces efficiency. They also struggle to maintain accuracy in complex multi-image, multi-concept settings. These limitations restrict the broader deployment of LVLM-based systems. Therefore, this paper proposes in-context prompt tuning (ICPT). Specifically, ICPT employs a lightweight projection module capable of operating in complex scenarios to extract fine-grained visual semantics from multiple reference images, seamlessly transforming these features alongside identity-label mappings into continuous prompts. To maximize computational efficiency, this module adaptively determines the prompt length based on the intrinsic visual complexity of each concept. Crucially, to overcome the environmental biases and cross-concept interference prevalent in real-world applications, we introduce two novel geometric regularizations. These constraints refine prompt representations by decoupling key identities from transient environmental states and separating concepts to avoid semantic confusion. Extensive experiments show that ICPT achieves state-of-the-art personalization accuracy across diverse tasks and LVLM backbones.

Zhaochen Su, Jun Zhang, Xiaoye Qu et al.

Large language models (LLMs) have achieved impressive advancements across numerous disciplines, yet the critical issue of knowledge conflicts, a major source of hallucinations, has rarely been studied. Only a few research explored the conflicts between the inherent knowledge of LLMs and the retrieved contextual knowledge. However, a thorough assessment of knowledge conflict in LLMs is still missing. Motivated by this research gap, we present ConflictBank, the first comprehensive benchmark developed to systematically evaluate knowledge conflicts from three aspects: (i) conflicts encountered in retrieved knowledge, (ii) conflicts within the models' encoded knowledge, and (iii) the interplay between these conflict forms. Our investigation delves into four model families and twelve LLM instances, meticulously analyzing conflicts stemming from misinformation, temporal discrepancies, and semantic divergences. Based on our proposed novel construction framework, we create 7,453,853 claim-evidence pairs and 553,117 QA pairs. We present numerous findings on model scale, conflict causes, and conflict types. We hope our ConflictBank benchmark will help the community better understand model behavior in conflicts and develop more reliable LLMs.

Qingyu Yin, Yulun Wu, Zhennan Shen et al.

We systematically evaluate Parameter-Efficient Fine-Tuning (PEFT) methods under the paradigm of Reinforcement Learning with Verifiable Rewards (RLVR). RLVR incentivizes language models to enhance their reasoning capabilities through verifiable feedback; however, while methods like LoRA are commonly used, the optimal PEFT architecture for RLVR remains unidentified. In this work, we conduct the first comprehensive evaluation of over 12 PEFT methodologies across the DeepSeek-R1-Distill families on mathematical reasoning benchmarks. Our empirical results challenge the default adoption of standard LoRA with three main findings. First, we demonstrate that structural variants, such as DoRA, AdaLoRA, and MiSS, consistently outperform LoRA. Second, we uncover a spectral collapse phenomenon in SVD-informed initialization strategies (\textit{e.g.,} PiSSA, MiLoRA), attributing their failure to a fundamental misalignment between principal-component updates and RL optimization. Furthermore, our ablations reveal that extreme parameter reduction (\textit{e.g.,} VeRA, Rank-1) severely bottlenecks reasoning capacity. We further conduct ablation studies and scaling experiments to validate our findings. This work provides a definitive guide for advocating for more exploration for parameter-efficient RL methods.

Siheng Wang, Yanshu Li, Bohan Hu et al.

Open-vocabulary Object Detection (OVOD) enables models to recognize objects beyond predefined categories, but existing approaches remain limited in practical deployment. On the one hand, multimodal designs often incur substantial computational overhead due to their reliance on text encoders at inference time. On the other hand, tightly coupled training objectives introduce a trade-off between closed-set detection accuracy and open-world generalization. Thus, we propose Decoupled Cognition DETR (DeCo-DETR), a vision-centric framework that addresses these challenges through a unified decoupling paradigm. Instead of depending on online text encoding, DeCo-DETR constructs a hierarchical semantic prototype space from region-level descriptions generated by pre-trained LVLMs and aligned via CLIP, enabling efficient and reusable semantic representation. Building upon this representation, the framework further disentangles semantic reasoning from localization through a decoupled training strategy, which separates alignment and detection into parallel optimization streams. Extensive experiments on standard OVOD benchmarks demonstrate that DeCo-DETR achieves competitive zero-shot detection performance while significantly improving inference efficiency. These results highlight the effectiveness of decoupling semantic cognition from detection, offering a practical direction for scalable OVOD systems.

Siheng Wang, Zhengdao Li, Yanshu Li et al.

Object detection has advanced significantly in the closed-set setting, but real-world deployment remains limited by two challenges: poor generalization to unseen categories and insufficient robustness under adverse conditions. Prior research has explored these issues separately: visible-infrared detection improves robustness but lacks generalization, while open-world detection leverages vision-language alignment strategy for category diversity but struggles under extreme environments. This trade-off leaves robustness and diversity difficult to achieve simultaneously. To mitigate these issues, we propose \textbf{C3-OWD}, a curriculum cross-modal contrastive learning framework that unifies both strengths. Stage~1 enhances robustness by pretraining with RGBT data, while Stage~2 improves generalization via vision-language alignment. To prevent catastrophic forgetting between two stages, we introduce an Exponential Moving Average (EMA) mechanism that theoretically guarantees preservation of pre-stage performance with bounded parameter lag and function consistency. Experiments on FLIR, OV-COCO, and OV-LVIS demonstrate the effectiveness of our approach: C3-OWD achieves $80.1$ AP$^{50}$ on FLIR, $48.6$ AP$^{50}_{\text{Novel}}$ on OV-COCO, and $35.7$ mAP$_r$ on OV-LVIS, establishing competitive performance across both robustness and diversity evaluations. Code available at: https://github.com/justin-herry/C3-OWD.git.

Xi Xiao, Chenrui Ma, Yunbei Zhang et al.

Low-Rank Adaptation (LoRA) has become a cornerstone of parameter-efficient fine-tuning (PEFT). Yet, its efficacy is hampered by two fundamental limitations: semantic drift, by treating all update directions with equal importance, and structural incoherence, from adapting layers independently, resulting in suboptimal, uncoordinated updates. To remedy these, we propose StructLoRA, a framework that addresses both limitations through a principled, dual-component design: (1) an Information Bottleneck-guided filter that prunes task-irrelevant directions to mitigate semantic drift, and (2) a lightweight, training-only graph-based coordinator that enforces inter-layer consistency to resolve structural incoherence. Extensive experiments across large language model , vision language model, and vision model (including LLaMA, LLaVA, and ViT) demonstrate that StructLoRA consistently establishes a new state-of-the-art, outperforming not only vanilla LoRA but also advanced dynamic rank allocation and sparsity-based methods. Notably, the benefits are particularly pronounced in challenging low-rank and low-data regimes. Crucially, since our proposed modules operate only during training, StructLoRA enhances performance with zero additional inference cost, advancing the focus of PEFT -- from mere parameter compression to a more holistic optimization of information quality and structural integrity.

Xi Xiao, Zhuxuanzi Wang, Mingqiao Mo et al.

The deployment of automated pavement defect detection is often hindered by poor cross-domain generalization. Supervised detectors achieve strong in-domain accuracy but require costly re-annotation for new environments, while standard self-supervised methods capture generic features and remain vulnerable to domain shift. We propose \ours, a self-supervised framework that \emph{visually probes} target domains without labels. \ours introduces a Self-supervised Prompt Enhancement Module (SPEM), which derives defect-aware prompts from unlabeled target data to guide a frozen ViT backbone, and a Domain-Aware Prompt Alignment (DAPA) objective, which aligns prompt-conditioned source and target representations. Experiments on four challenging benchmarks show that \ours consistently outperforms strong supervised, self-supervised, and adaptation baselines, achieving robust zero-shot transfer, improved resilience to domain variations, and high data efficiency in few-shot adaptation. These results highlight self-supervised prompting as a practical direction for building scalable and adaptive visual inspection systems. Source code is publicly available: https://github.com/xixiaouab/PROBE/tree/main

Siheng Wan, Zhengtao Yao, Zhengdao Li et al.

Modern Text-to-Image (T2I) generation increasingly relies on token-centric architectures that are trained with self-supervision, yet effectively fusing text with visual tokens remains a challenge. We propose \textbf{JEPA-T}, a unified multimodal framework that encodes images and captions into discrete visual and textual tokens, processed by a joint-embedding predictive Transformer. To enhance fusion, we incorporate cross-attention after the feature predictor for conditional denoising while maintaining a task-agnostic backbone. Additionally, raw texts embeddings are injected prior to the flow matching loss to improve alignment during training. During inference, the same network performs both class-conditional and free-text image generation by iteratively denoising visual tokens conditioned on text. Evaluations on ImageNet-1K demonstrate that JEPA-T achieves strong data efficiency, open-vocabulary generalization, and consistently outperforms non-fusion and late-fusion baselines. Our approach shows that late architectural fusion combined with objective-level alignment offers an effective balance between conditioning strength and backbone generality in token-based T2I.The code is now available: https://github.com/justin-herry/JEPA-T.git

Wenquan Lu, Yuechuan Yang, Kyle Lee et al.

Chain-of-thought (CoT) reasoning has enabled transformer-based language models to excel at complex mathematics and multi-step planning. However, in standard decoder-only architectures, these reasoning steps are externalized in natural language, improving interpretability at the cost of efficiency. To capture reasoning that is not easily represented in words, many works have explored recurrent architectures that aim to internalize reasoning in latent space, potentially supporting latent CoT. In this paper, we investigate whether such reasoning structures emerge in Huginn-3.5B, a depth-recurrent Transformer that reuses layers at inference time without increasing parameter count. We examine the model's internal behavior on arithmetic tasks using a suite of probing techniques including the Logit Lens and Coda Lens. Our findings reveal limited evidence of interpretable latent CoT by tracking rank trajectories of final and intermediate result tokens. Furthermore, we uncover significant probing inconsistencies across recurrent blocks, where the interpretability of hidden states depends heavily on both the layer index and the decoding method. Finally, we empirically show that increasing recurrence depth yields only marginal gains and falls well short of models that explicitly externalize reasoning steps. The code is available at https://github.com/wenquanlu/huginn-latent-cot.

Zhaochen Su, Peng Xia, Hangyu Guo et al.

Recent progress in multimodal reasoning has been significantly advanced by textual Chain-of-Thought (CoT), a paradigm where models conduct reasoning within language. This text-centric approach, however, treats vision as a static, initial context, creating a fundamental "semantic gap" between rich perceptual data and discrete symbolic thought. Human cognition often transcends language, utilizing vision as a dynamic mental sketchpad. A similar evolution is now unfolding in AI, marking a fundamental paradigm shift from models that merely think about images to those that can truly think with images. This emerging paradigm is characterized by models leveraging visual information as intermediate steps in their thought process, transforming vision from a passive input into a dynamic, manipulable cognitive workspace. In this survey, we chart this evolution of intelligence along a trajectory of increasing cognitive autonomy, which unfolds across three key stages: from external tool exploration, through programmatic manipulation, to intrinsic imagination. To structure this rapidly evolving field, our survey makes four key contributions. (1) We establish the foundational principles of the think with image paradigm and its three-stage framework. (2) We provide a comprehensive review of the core methods that characterize each stage of this roadmap. (3) We analyze the critical landscape of evaluation benchmarks and transformative applications. (4) We identify significant challenges and outline promising future directions. By providing this structured overview, we aim to offer a clear roadmap for future research towards more powerful and human-aligned multimodal AI.

Weihai Lu, Zhejun Zhao, Yanshu Li et al.

Multimodal Stance Detection (MSD) is crucial for understanding public discourse, yet effectively fusing text and image, especially with conflicting signals, remains challenging. Existing methods often face difficulties with contextual grounding, cross-modal interpretation ambiguity, and single-pass reasoning fragility. To address these, we propose Retrieval-Augmented Multi-modal Multi-agent Stance Detection (MM-StanceDet), a novel multi-agent framework integrating Retrieval Augmentation for contextual grounding, specialized Multimodal Analysis agents for nuanced interpretation, a Reasoning-Enhanced Debate stage for exploring perspectives, and Self-Reflection for robust adjudication. Extensive experiments on five datasets demonstrate MM-StanceDet significantly outperforms state-of-the-art baselines, validating the efficacy of its multi-agent architecture and structured reasoning stages in addressing complex multimodal stance challenges.

Yanshu Li, Jianjiang Yang, Tian Yun et al.

Multimodal in-context learning (ICL) has emerged as a key mechanism for harnessing the capabilities of large vision-language models (LVLMs). However, its effectiveness remains highly sensitive to the quality of input ICL sequences, particularly for tasks involving complex reasoning or open-ended generation. A major limitation is our limited understanding of how LVLMs actually exploit these sequences during inference. To bridge this gap, we systematically interpret multimodal ICL through the lens of task mapping, which reveals how local and global relationships within and among demonstrations guide model reasoning. Building on this insight, we present TACO, a lightweight transformer-based model equipped with task-aware attention that dynamically configures ICL sequences. By injecting task-mapping signals into the autoregressive decoding process, TACO creates a bidirectional synergy between sequence construction and task reasoning. Experiments on five LVLMs and nine datasets demonstrate that TACO consistently surpasses baselines across diverse ICL tasks. These results position task mapping as a novel and valuable perspective for interpreting and improving multimodal ICL.

Minjie Shen, Yanshu Li, Lulu Chen et al.

Manus AI is a general-purpose AI agent introduced in early 2025, marking a significant advancement in autonomous artificial intelligence. Developed by the Chinese startup Monica.im, Manus is designed to bridge the gap between "mind" and "hand" - combining the reasoning and planning capabilities of large language models with the ability to execute complex, end-to-end tasks that produce tangible outcomes. This paper presents a comprehensive overview of Manus AI, exploring its core technical architecture, diverse applications across sectors such as healthcare, finance, manufacturing, robotics, and gaming, as well as its key strengths, current limitations, and future potential. Positioned as a preview of what lies ahead, Manus AI represents a shift toward intelligent agents that can translate high-level intentions into real-world actions, heralding a new era of human-AI collaboration.

Yanshu Li, Yi Cao, Hongyang He et al.

Multimodal in-context learning (ICL) equips Large Vision-language Models (LVLMs) with the ability to adapt to new tasks via multiple user-provided demonstrations, without requiring any model parameter updates. However, its effectiveness is constrained by the token-intensive nature of multimodal inputs and the complexity of cross-modal few-shot reasoning, which together hinder LVLMs from extracting useful patterns from demonstrations. To address these challenges, we propose \textbf{M$^2$IV}, a novel representation engineering approach that replaces explicit token-level demonstrations with a set of learnable Multimodal In-context Vectors directly injected into the residual streams of LVLMs. By analyzing the distinct roles of multi-head attention (MHA) and multi-layer perceptrons (MLP) in the ICL process, we design a training strategy that enables M$^2$IV to perform fine-grained semantic distillation and robust cross-modal representation learning. M$^2$IV not only improves performance across diverse tasks and LVLMs but also significantly reduces token overhead, enabling graceful scaling to many-shot scenarios. To further enhance usability, we introduce \textbf{VLibrary}, a repository that stores trained M$^2$IVs for flexible retrieval and injection. With VLibrary, users can steer pre-trained LVLMs in a customized manner that meets diverse requirements. Extensive experiments demonstrate that M$^2$IV consistently outperforms vanilla ICL and prior representation engineering baselines, achieving an average accuracy gain of 3.74\% with substantial improvements in overall efficiency.

Yanshu Li, Jianjiang Yang, Ziteng Yang et al.

Multimodal in-context learning (ICL) is emerging as a key capability that enables large vision-language models (LVLMs) to adapt to novel tasks without parameter updates, expanding their utility across various real-world applications. However, ICL remains unstable, even with well-matched in-context demonstrations (ICDs), suggesting that LVLMs struggle to fully utilize the provided context. While existing efforts focus on prompt engineering or post-hoc logit calibration, we instead investigate the underlying attention dynamics to overcome LVLMs' inherent limitations. We identify two critical deficits in their self-attention that impair effective ICL. To bridge the gap, we propose \textbf{Context-Aware Modulated Attention} (CAMA), a plug-and-play and training-free method that dynamically modulates LVLM's attention logits based on the input in-context sequence. CAMA employs a two-stage attention modulation to address both identified deficits, enhancing the focus on semantically significant tokens, particularly visual ones. Across four LVLMs and seven benchmarks, CAMA consistently outperforms vanilla models and baselines, demonstrating great effectiveness and generalization. It can also activate the desired effects of prompt engineering methods and remains robust under diverse sequence configurations. Thus, CAMA paves the way for deeper explorations of attention dynamics to advance multimodal reasoning.

Jiaqian Li, Yanshu Li, Kuan-Hao Huang

Steering vectors (SVs) offer a lightweight way to control large language models (LLMs) at inference time by shifting hidden activations, providing a practical middle ground between prompting and fine-tuning. Yet SVs can be unreliable in practice. Some concepts are unsteerable, and even when steering helps on average it can backfire for a non-trivial fraction of inputs. Reliability also degrades in long-form generation and multi-attribute steering. We take a geometric view of these failures. A static SV applies the same update vector everywhere in representation space, implicitly assuming that the concept-improving direction is constant across contexts. When the locally effective direction varies with the current activation, a single global vector can become misaligned, which yields weak or reversed effects. Guided by this perspective, we propose Steering Vector Fields (SVF), which learns a differentiable concept scoring function whose local gradient defines the steering direction at each activation, making interventions explicitly context-dependent. This formulation supports coordinated multi-layer interventions in a shared, aligned concept space, and enables efficient long-form and multi-attribute control within a unified framework. Across multiple LLMs and steering tasks, SVF delivers stronger and more reliable control, improving the practicality of inference-time steering.

Zhiqing Cui, Jiahao Yuan, Hanqing Wang et al.

Scientific diagrams are vital tools for communicating structured knowledge across disciplines. However, they are often published as static raster images, losing symbolic semantics and limiting reuse. While Multimodal Large Language Models (MLLMs) offer a pathway to bridging vision and structure, existing methods lack semantic control and structural interpretability, especially on complex diagrams. We propose Draw with Thought (DwT), a training-free framework that guides MLLMs to reconstruct diagrams into editable mxGraph XML code through cognitively-grounded Chain-of-Thought reasoning. DwT enables interpretable and controllable outputs without model fine-tuning by dividing the task into two stages: Coarse-to-Fine Planning, which handles perceptual structuring and semantic specification, and Structure-Aware Code Generation, enhanced by format-guided refinement. To support evaluation, we release Plot2XML, a benchmark of 247 real-world scientific diagrams with gold-standard XML annotations. Extensive experiments across eight MLLMs show that our approach yields high-fidelity, semantically aligned, and structurally valid reconstructions, with human evaluations confirming strong alignment in both accuracy and visual aesthetics, offering a scalable solution for converting static visuals into executable representations and advancing machine understanding of scientific graphics.

Yanshu Li, Jianjiang Yang, Zhennan Shen et al.

Modern large vision-language models (LVLMs) convert each input image into a large set of tokens, far outnumbering the text tokens. Although this improves visual perception, it introduces severe image token redundancy. Because image tokens carry sparse information, many add little to reasoning, yet greatly increase inference cost. The emerging image token pruning methods tackle this issue by identifying the most important tokens and discarding the rest. These methods can raise efficiency with only modest performance loss. However, most of them only consider single-image tasks and overlook multimodal in-context learning (ICL), where redundancy is greater and efficiency is more critical. Redundant tokens weaken the advantage of multimodal ICL for rapid domain adaptation and cause unstable performance. Applying existing pruning methods in this setting leads to large accuracy drops, exposing a clear gap and the need for new techniques. Thus, we propose Contextually Adaptive Token Pruning (CATP), a training-free pruning method targeted at multimodal ICL. CATP consists of two stages that perform progressive pruning to fully account for the complex cross-modal interactions in the input sequence. After removing 77.8\% of the image tokens, CATP produces an average performance gain of 0.6\% over the vanilla model on four LVLMs and eight benchmarks, exceeding all baselines remarkably. Meanwhile, it effectively improves efficiency by achieving an average reduction of 10.78\% in inference latency. CATP enhances the practical value of multimodal ICL and lays the groundwork for future progress in interleaved image-text scenarios.

Yanshu Li

Multimodal in-context learning (ICL) has emerged as a key capability of Large Vision-Language Models (LVLMs), driven by their increasing scale and applicability. Despite its promise, effective ICL in the multimodal setting remains challenging due to the inherent complexity of image-text inputs and the high sensitivity of ICL performance to input configurations. In this work, we shed light on the core mechanism underlying multimodal ICL, identifying task mapping as a crucial factor in configuring robust in-context demonstration (ICD) sequences. Building on these insights, we propose \textit{SabER}, a lightweight yet powerful decoder-only transformer equipped with task-aware attention, which intelligently selects and arranges ICDs from a demonstration library in an autoregressive fashion. This design enables fine-grained feature extraction and cross-modal reasoning, iteratively refining task mapping to generate high-quality ICD sequences. Through extensive experiments covering five LVLMs and nine benchmark datasets, SabER not only demonstrates strong empirical performance, but also provides deeper understanding of how task semantics interact with multimodal ICDs. Our findings highlight the importance of principled ICD sequence configuration and open new avenues to enhance multimodal ICL in a wide range of real-world scenarios.

Chengyan Wu, Bolei Ma, Yihong Liu et al.

Aspect-based sentiment analysis (ABSA) is a crucial task in information extraction and sentiment analysis, aiming to identify aspects with associated sentiment elements in text. However, existing ABSA datasets are predominantly English-centric, limiting the scope for multilingual evaluation and research. To bridge this gap, we present M-ABSA, a comprehensive dataset spanning 7 domains and 21 languages, making it the most extensive multilingual parallel dataset for ABSA to date. Our primary focus is on triplet extraction, which involves identifying aspect terms, aspect categories, and sentiment polarities. The dataset is constructed through an automatic translation process with human review to ensure quality. We perform extensive experiments using various baselines to assess performance and compatibility on M-ABSA. Our empirical findings highlight that the dataset enables diverse evaluation tasks, such as multilingual and multi-domain transfer learning, and large language model evaluation, underscoring its inclusivity and its potential to drive advancements in multilingual ABSA research.

Xiaodong Qu, Andrews Damoah, Joshua Sherwood et al.

Artificial Intelligence (AI) agents have rapidly evolved from specialized, rule-based programs to versatile, learning-driven autonomous systems capable of perception, reasoning, and action in complex environments. The explosion of data, advances in deep learning, reinforcement learning, and multi-agent coordination have accelerated this transformation. Yet, designing and deploying unified AI agents that seamlessly integrate cognition, planning, and interaction remains a grand challenge. In this review, we systematically examine the architectural principles, foundational components, and emergent paradigms that define the landscape of contemporary AI agents. We synthesize insights from cognitive science-inspired models, hierarchical reinforcement learning frameworks, and large language model-based reasoning. Moreover, we discuss the pressing ethical, safety, and interpretability concerns associated with deploying these agents in real-world scenarios. By highlighting major breakthroughs, persistent challenges, and promising research directions, this review aims to guide the next generation of AI agent systems toward more robust, adaptable, and trustworthy autonomous intelligence.

He Zhu, Yanshu Li, Wen Liu et al.

Textual adversarial attacks pose a serious security threat to Natural Language Processing (NLP) systems by introducing imperceptible perturbations that mislead deep learning models. While adversarial example detection offers a lightweight alternative to robust training, existing methods typically rely on prior knowledge of attacks, white-box access to the victim model, or numerous queries, which severely limits their practical deployment. This paper introduces RTD-Guard, a novel black-box framework for detecting textual adversarial examples. Our key insight is that word-substitution perturbations in adversarial attacks closely resemble the "replaced tokens" that a Replaced Token Detection (RTD) discriminator is pre-trained to identify. Leveraging this, RTD-Guard employs an off-the-shelf RTD discriminator-without fine-tuning-to localize suspicious tokens, masks them, and detects adversarial examples by observing the prediction confidence shift of the victim model before and after intervention. The entire process requires no adversarial data, model tuning, or internal model access, and uses only two black-box queries. Comprehensive experiments on multiple benchmark datasets demonstrate that RTD-Guard effectively detects adversarial texts generated by diverse state-of-the-art attack methods. It surpasses existing detection baselines across multiple metrics, offering a highly efficient, practical, and resource-light defense mechanism-particularly suited for real-world deployment in resource-constrained or privacy-sensitive environments.

Rui Wu, Yihao Quan, Zeru Shi et al.

Safety-aligned Large Language Models (LLMs) still show two dominant failure modes: they are easily jailbroken, or they over-refuse harmless inputs that contain sensitive surface signals. We trace both to a common cause: current models reason weakly about links between actions and outcomes and over-rely on surface-form signals, lexical or stylistic cues that do not encode consequences. We define this failure mode as Consequence-blindness. To study consequence-blindness, we build a benchmark named CB-Bench covering four risk scenarios that vary whether semantic risk aligns with outcome risk, enabling evaluation under both matched and mismatched conditions which are often ignored by existing safety benchmarks. Mainstream models consistently fail to separate these risks and exhibit consequence-blindness, indicating that consequence-blindness is widespread and systematic. To mitigate consequence-blindness, we introduce CS-Chain-4k, a consequence-reasoning dataset for safety alignment. Models fine-tuned on CS-Chain-4k show clear gains against semantic-camouflage jailbreaks and reduce over-refusal on harmless inputs, while maintaining utility and generalization on other benchmarks. These results clarify the limits of current alignment, establish consequence-aware reasoning as a core alignment goal and provide a more practical and reproducible evaluation path.

Jiaqian Li, Yanshu Li, Ligong Han et al.

Implicit in-context learning (ICL) has newly emerged as a promising paradigm that simulates ICL behaviors in the representation space of Large Language Models (LLMs), aiming to attain few-shot performance at zero-shot cost. However, existing approaches largely rely on injecting shift vectors into residual flows, which are typically constructed from labeled demonstrations or task-specific alignment. Such designs fall short of utilizing the structural mechanisms underlying ICL and suffer from limited generalizability. To address this, we propose In-Context Routing (ICR), a novel implicit ICL method that internalizes generalizable ICL patterns at the attention logits level. It extracts reusable structural directions that emerge during ICL and employs a learnable input-conditioned router to modulate attention logits accordingly, enabling a train-once-and-reuse framework. We evaluate ICR on 12 real-world datasets spanning diverse domains and multiple LLMs. The results show that ICR consistently outperforms prior implicit ICL methods that require task-specific retrieval or training, while demonstrating robust generalization to out-of-domain tasks where existing methods struggle. These findings position ICR to push the boundary of ICL's practical value.

Hongyang He, Xinyuan Song, Yangfan He et al.

We introduce TRiCo, a novel triadic game-theoretic co-training framework that rethinks the structure of semi-supervised learning by incorporating a teacher, two students, and an adversarial generator into a unified training paradigm. Unlike existing co-training or teacher-student approaches, TRiCo formulates SSL as a structured interaction among three roles: (i) two student classifiers trained on frozen, complementary representations, (ii) a meta-learned teacher that adaptively regulates pseudo-label selection and loss balancing via validation-based feedback, and (iii) a non-parametric generator that perturbs embeddings to uncover decision boundary weaknesses. Pseudo-labels are selected based on mutual information rather than confidence, providing a more robust measure of epistemic uncertainty. This triadic interaction is formalized as a Stackelberg game, where the teacher leads strategy optimization and students follow under adversarial perturbations. By addressing key limitations in existing SSL frameworks, such as static view interactions, unreliable pseudo-labels, and lack of hard sample modeling, TRiCo provides a principled and generalizable solution. Extensive experiments on CIFAR-10, SVHN, STL-10, and ImageNet demonstrate that TRiCo consistently achieves state-of-the-art performance in low-label regimes, while remaining architecture-agnostic and compatible with frozen vision backbones.

Jianjiang Yang, Ziyan Huang, Yanshu li et al.

Despite remarkable progress in image quality and prompt fidelity, text-to-image (T2I) diffusion models continue to exhibit persistent "hallucinations", where generated content subtly or significantly diverges from the intended prompt semantics. While often regarded as unpredictable artifacts, we argue that these failures reflect deeper, structured misalignments within the generative process. In this work, we propose a cognitively inspired perspective that reinterprets hallucinations as trajectory drift within a latent alignment space. Empirical observations reveal that generation unfolds within a multiaxial cognitive tension field, where the model must continuously negotiate competing demands across three key critical axes: semantic coherence, structural alignment, and knowledge grounding. We then formalize this three-axis space as the Hallucination Tri-Space and introduce the Alignment Risk Code (ARC): a dynamic vector representation that quantifies real-time alignment tension during generation. The magnitude of ARC captures overall misalignment, its direction identifies the dominant failure axis, and its imbalance reflects tension asymmetry. Based on this formulation, we develop the TensionModulator (TM-ARC): a lightweight controller that operates entirely in latent space. TM-ARC monitors ARC signals and applies targeted, axis-specific interventions during the sampling process. Extensive experiments on standard T2I benchmarks demonstrate that our approach significantly reduces hallucination without compromising image quality or diversity. This framework offers a unified and interpretable approach for understanding and mitigating generative failures in diffusion-based T2I systems.

Jianjiang Yang, Yanshu li, Ziyan Huang

While Multimodal Large Language Models (MLLMs) have achieved remarkable progress in open-ended visual question answering, they remain vulnerable to hallucinations. These are outputs that contradict or misrepresent input semantics, posing a critical challenge to the reliability and factual consistency. Existing methods often rely on external verification or post-hoc correction, lacking an internal mechanism to validate outputs directly during training. To bridge this gap, we propose ReLoop, a unified closed-loop training framework that encourages multimodal consistency for cross-modal understanding in MLLMs. ReLoop adopts a ring-shaped structure that integrates three complementary consistency feedback mechanisms, obliging MLLMs to "seeing twice and thinking backwards". Specifically, ReLoop employs the frozen Consistency Feedback Plugin (CFP), comprising semantic reconstruction, visual description, and an attention supervision module for attention alignment. These components collectively enforce semantic reversibility, visual consistency, and interpretable attention, enabling the model to correct its outputs during training. Extensive evaluations and analyses demonstrate the effectiveness of ReLoop in reducing hallucination rates across multiple benchmarks, establishing a robust method for hallucination mitigation in MLLMs. We will release our source code and data in the camera-ready version.

Ziteng Yang, Jingzehua Xu, Yanshu Li et al.

Zero-shot anomaly detection (ZSAD) aims to detect anomalies without any target domain training samples, relying solely on external auxiliary data. Existing CLIP-based methods attempt to activate the model's ZSAD potential via handcrafted or static learnable prompts. The former incur high engineering costs and limited semantic coverage, whereas the latter apply identical descriptions across diverse anomaly types, thus fail to adapt to complex variations. Furthermore, since CLIP is originally pretrained on large-scale classification tasks, its anomaly segmentation quality is highly sensitive to the exact wording of class names, severely constraining prompting strategies that depend on class labels. To address these challenges, we introduce ViP$^{2}$-CLIP. The key insight of ViP$^{2}$-CLIP is a Visual-Perception Prompting (ViP-Prompt) mechanism, which fuses global and multi-scale local visual context to adaptively generate fine-grained textual prompts, eliminating manual templates and class-name priors. This design enables our model to focus on precise abnormal regions, making it particularly valuable when category labels are ambiguous or privacy-constrained. Extensive experiments on 15 industrial and medical benchmarks demonstrate that ViP$^{2}$-CLIP achieves state-of-the-art performance and robust cross-domain generalization.