Search arXiv⌕ Search

arXiv subjects

Yuchen Zhang

Publications and source records attributed to Yuchen Zhang.

At least 73 records · Page 4Linked to original sources
arXiv2026

MT-JailBench: A Modular Benchmark for Understanding Multi-Turn Jailbreak Attacks

Multi-turn jailbreaks exploit the ability of large language models to accumulate and act on conversational context. Instead of stating a harmful request directly, an attacker can gradually steer the conversation toward an unsafe answer. Recent methods demonstrate this risk, but they are usually evaluated as black-box pipelines with different budgets, judges, retry rules, and strategy generation procedures. As a result, it is often unclear whether reported gains reflect stronger attack mechanisms or different experimental conditions. We introduce MT-JailBench, a modular evaluation framework for benchmarking multi-turn jailbreaks under fixed conditions. MT-JailBench implements each attack as five interacting modules: evaluation function, attack strategy, prompt generation, prompt refinement, and flow control. This design enables fair comparison across attack methods and component-wise analysis of what drives attack success. Using MT-JailBench, we find that resource budgets and evaluation functions are major confounders: controlling turns, retries, interactions, sampled strategies, and judges substantially change the ranking of attacks. At the component level, prompt generation accounts for most performance variation, while refinement and flow control provide moderate gains. We also find that explicit dynamic strategy generation is not always necessary; stochastic sampling from a fixed strategy can rival more elaborate diversification mechanisms. Finally, recomposing the best components yields a strong attack configuration that outperforms its source attacks and generalizes across diverse target LLMs. MT-JailBench therefore provides a modular framework for comparing multi-turn jailbreaks, understanding the impact of components, and guiding stronger red-teaming evaluations.

cs.CR↗
arXiv2026

Semantic Communication for Multi-Satellite Massive MIMO Transmission: A Mixture of Cooperative Modes Framework

This paper investigates semantic communications (SemComs) for multi-satellite cooperative massive multiple-input multiple-output (MIMO) transmission, where multiple massive-MIMO satellites jointly serve a common set of multi-antenna user terminals. For the first time, SemComs with image transmission task are integrated into satellite massive MIMO and multi-satellite cooperative transmission. For the two representative cooperative modes, namely coherent transmission (CT) and non-coherent transmission (NCT), we develop multi-satellite CT (MSCT) and multi-satellite NCT (MSNCT) SemCom frameworks, respectively. MSCT adopts a symmetric architecture, whereas MSNCT introduces transmitter-side stream allocation and a two-stage receiver design that combines per-stream semantic extraction with cross-stream semantic-interference exploitation. To instantiate MSCT, we further design a symmetric encoder and decoder network based on hybrid Swin-Transformer and lightweight bottleneck convolutional neural network (CNN) blocks, termed HSTC, where Swin Transformer provides scalable computation and the CNN branch improves performance and convergence. For MSNCT, a Transformer-based backbone is employed to support cross-stream interference exploitation through global attention. Building on these two frameworks, we propose a mixture of cooperative modes (MoCM) framework, in which a permutation-invariant network dynamically switches between MSCT and MSNCT using multi-satellite statistical channel state information, thereby balancing semantic performance and complexity. Simulation results under practical configurations demonstrate the performance gains of the proposed frameworks.

eess.SP↗
arXiv2026

Implicit Compression Regularization: Concise Reasoning via Internal Shorter Distributions in RL Post-Training

Reinforcement learning with verifiable rewards improves LLM reasoning but often induces overthinking, where models generate unnecessarily long reasoning traces. Existing methods mainly rely on length penalties or early-exit strategies; however, the former may degrade accuracy and induce underthinking, whereas the latter assumes that substantial portions of reasoning traces can be safely truncated. To obtain a compression signal without these limitations, we revisit the training dynamics of existing compression methods. We observe that the length--accuracy correlation is initially negative but continually increases during compression, indicating that shorter responses are initially more likely to be correct but gradually lose this property as the policy moves toward underthinking. Based on this observation, we formalize overthinking: a negative correlation indicates an overthinking regime, while a positive one indicates underthinking. When overthinking, the shortest correct responses are shorter than the group-average response length in expectation, making them natural compression targets already present in on-policy rollouts. We therefore propose \emph{Implicit Compression Regularization} (ICR), an on-policy regularization method whose compression signal comes from a virtual shorter distribution induced by the shortest correct responses in rollout groups, guiding the policy toward concise yet correct trajectories. Training dynamics show that ICR maintains a better length--accuracy correlation during compression, indicating that short responses remain better aligned with correctness instead of drifting toward underthinking. Experiments on three reasoning backbones and multiple mathematical and knowledge-intensive benchmarks show that ICR consistently shortens responses while preserving or improving accuracy, achieving a stronger accuracy--length Pareto frontier.

cs.AI↗
arXiv2026

GESR: Graph-Based Edge Semantic Reconstruction for Stealthy Communication Detection with Benign-Only Training

Detecting stealthy malicious communications from flow logs under benign-only training remains a critical challenge in network security. Malicious communications often camouflage as normal traffic like standard HTTPS flows. Conventional intrusion detectors rely strictly on known labeled attacks. Alternatively, they score flows completely independently. These approaches fail against sparse and context-dependent suspicious activity. To capture this essential context, graph anomaly detectors have been introduced to add valuable relational information to the analysis. However, existing methods fail to test the structural consistency of specific communication edges. To overcome these fundamental limitations, we present GESR, a novel graph-based framework for detecting suspicious communications and anomalous hosts under a benign-only training setting. GESR models complex network activity as attributed communication graphs. It cleverly reconstructs edge semantics entirely from local structural context rather than isolated features. This non-intuitive design forces the framework to predict expected communication patterns from neighborhood topologies. Attackers cannot easily manipulate this deep structural dependency. The model then converts the resulting structural inconsistencies into host-level anomaly scores. It utilizes robust Median Absolute Deviation (MAD) calibration for this final step. We evaluate GESR extensively on CTU-13 and CICIDS2017 datasets. These evaluations strictly impose tight false-positive operating constraints. On CICIDS2017, GESR achieves an outstanding ROC-AUC of 0.9753. It also yields a high TPR of 0.8569 at a strict 5% FPR threshold. GESR consistently outperforms existing methods across both evaluated benchmarks. The results prove that structure-conditioned edge reconstruction is a credible direction for practical intrusion detection.

cs.CR↗
arXiv2026

Teaching Thinking Models to Reason with Tools: A Full-Pipeline Recipe for Tool-Integrated Reasoning

Tool-integrated reasoning (TIR) offers a direct way to extend thinking models beyond the limits of text-only reasoning. Paradoxically, we observe that tool-enabled evaluation can degrade reasoning performance even when the strong thinking models make almost no actual tool calls. In this paper, we investigate how to inject natural tool-use behavior into a strong thinking model without sacrificing its no-tool reasoning ability, and present a comprehensive TIR recipe. We highlight that (i) the effectiveness of TIR supervised fine-tuning (SFT) hinges on the learnability of teacher trajectories, which should prioritize problems inherently suited for tool-augmented solutions; (ii) controlling the proportion of tool-use trajectories could mitigate the catastrophic forgetting of text-only reasoning capacity; (iii) optimizing for pass@k and response length instead of training loss could maximize TIR SFT gains while preserving headroom for reinforcement learning (RL) exploration; (iv) a stable RL with verifiable rewards (RLVR) stage, built upon suitable SFT initialization and explicit safeguards against mode collapse, provides a simple yet remarkably effective solution. When applied to Qwen3 thinking models at 4B and 30B scales, our recipe yields models that achieve state-of-the-art performance in a wide range of benchmarks among open-source models, such as 96.7% and 99.2% on AIME 2025 for 4B and 30B, respectively.

cs.CL↗
arXiv2026

Beyond Known Objects: A Novel Framework for Open-Set Object Detection using Negative-Aware Norm

Open-Set Object Detection (OSOD) is crucial for autonomous driving, where perception systems must recognize and localize both known and previously unseen objects in complex, dynamic environments. While recent approaches deliver promising results, they often require retraining the detector extensively to learn objectness, which describes the likelihood that a bounding box tightly encloses a valid object, regardless of whether its category was learned during training. Deviating from existing work, we hypothesize that standard off-the-shelf detectors may already contain helpful cues for objectness, owing to their training on numerous and diverse known categories. Building on this idea, we propose NAN-SPOT, a training-light framework that does not require to retrain the base object detector and estimates objectness by leveraging a hidden layer metric called Negative-Aware Norm (NAN), requiring only minutes of training on just hundreds of images. To support comprehensive evaluation, we introduce COCO-Open, an expanded version of the existing COCO-Mixed dataset, increasing unknown object annotations from 433 to 1853, making it the most exhaustively labeled dataset for OSOD to the best of our knowledge. Experimental results demonstrate that NAN-SPOT achieves even better performance on unknown object detection than methods requiring heavy training, without compromising performance on known objects. This efficiency and robustness make NAN-SPOT a promising step towards open-world perception in autonomous driving.

cs.CV↗
arXiv2026

Decentralized Cooperative Beamforming for Networked LEO Satellites with Statistical CSI

Inter-satellite-link-enabled low-Earth-orbit (LEO) satellite constellations are evolving toward networked architectures that support constellation-level cooperation, enabling multiple satellites to jointly serve user terminals through cooperative beamforming. While such cooperation can substantially enhance link budgets and achievable rates, its practical realization is challenged by the scalability limitations of centralized beamforming designs and the stringent computational and signaling constraints of large LEO constellations. This paper develops a fully decentralized cooperative beamforming framework for networked LEO satellite downlinks. Using an ergodic-rate-based formulation, we first derive a centralized weighted minimum mean squared error (WMMSE) solution as a performance benchmark. Building on this formulation, we propose a topology-agnostic decentralized beamforming algorithm by localizing the benchmark and exchanging a set of globally coupled variables whose dimensions are independent of the antenna number and enforcing consensus over arbitrary connected inter-satellite networks. The resulting algorithm admits fully parallel execution across satellites. To further enhance scalability, we eliminate the consensus-related auxiliary variables in closed form and derive a low-complexity per-satellite update rule that is optimal to local iteration and admits a quasi-closed-form solution via scalar line search. Simulation results show that the proposed decentralized schemes closely approach centralized performance under practical inter-satellite topologies, while significantly reducing computational complexity and signaling overhead, enabling scalable cooperative beamforming for large LEO constellations.

eess.SP↗
arXiv2026

Characterizing the Consistency of the Emergent Misalignment Persona

Fine-tuning large language models (LLMs) on narrowly misaligned data generalizes to broadly misaligned behavior, a phenomenon termed emergent misalignment (EM). While prior work has found a correlation between harmful behavior and self-assessment in emergently misaligned models, it remains unclear how consistent this correspondence is across tasks and whether it varies across fine-tuning domains. We characterize the consistency of the EM persona by fine-tuning Qwen 2.5 32B Instruct on six narrowly misaligned domains (e.g., insecure code, risky financial advice, bad medical advice) and administering experiments including harmfulness evaluation, self-assessment, choosing between two descriptions of AI systems, output recognition, and score prediction. Our results reveal two distinct patterns: coherent-persona models, in which harmful behavior and self-reported misalignment are coupled, and inverted-persona models, which produce harmful outputs while identifying as aligned AI systems. These findings reveal a more fine-grained picture of the effects of emergent misalignment, calling into question the consistency of the EM persona.

cs.AI↗
arXiv2026

MiMo-Embodied: X-Embodied Foundation Model Technical Report

We open-source MiMo-Embodied, the first cross-embodied foundation model to successfully integrate and achieve state-of-the-art performance in both Autonomous Driving and Embodied AI. MiMo-Embodied sets new records across 17 embodied AI benchmarks in Task Planning, Affordance Prediction and Spatial Understanding, while also excelling in 12 autonomous driving benchmarks across Environmental Perception, Status Prediction, and Driving Planning. Across these tasks, MiMo-Embodied significantly outperforms existing open-source, closed-source, and specialized baselines. Our results indicate that through multi-stage learning, curated data construction, and CoT/RL fine-tuning, these two domains exhibit strong positive transfer and mutually reinforce one another. We provide a detailed analysis of our model design and training methodologies to facilitate further research. Code and models are available at https://github.com/XiaomiMiMo/MiMo-Embodied.

cs.RO↗
arXiv2026

Tri-Hybrid Beamforming Design for ISAC Systems with Reconfigurable Antennas

Integrated Sensing and Communication (ISAC) systems require efficient beamforming architectures to jointly support communication and sensing functionalities. To reduce hardware overhead, Hybrid Beamforming (HBF) has been widely studied and shown to achieve performance close to fully digital beamforming under practical hardware constraints. As a promising evolution, Reconfigurable Antenna (RA) technologies have recently emerged to further enhance beamforming Degrees of Freedom (DoFs) by dynamically reconfiguring antenna Electromagnetic(EM) characteristics, yet their integration into ISAC systems remains largely unexplored. In this paper, we investigate an RA-assisted ISAC system and develop a decoupled Triple-Hybrid Beamforming (Tri-HBF) framework that alternatively optimizes digital, analog, and EM beamformers to maximize the communication rate and sensing Signal-to-Clutter-plus-NoiseRatio (SCNR). For both Single-user Single-target (SUST) and Multiple-user Multiple-target (MUMT) scenarios, we first transform the original fractional objectives into fraction-free ones via methods tailored to their respective structures. The resulting problems are then solved via alternating optimization over different variable blocks. Closed-form updates are derived for all variables except the EM beamforming subproblem in the MUMT scenario. To further reduce the complexity introduced by Semidefinite Relaxation (SDR) in EM beamforming, we propose a low-complexity iterative approach across antennas with closed-form updates. Simulation results demonstrate that the proposed scheme significantly outperforms benchmark designs with conventional omnidirectional and directional antennas, achievingalmost 100% improvement in spectrum efficiency and 62.5% reduction in antenna overhead, thereby unveiling the

eess.SP↗
arXiv2026

A Pair of Multiplication-Type Operators in Quaternionic Analysis and the 2-Cauchy-Fueter Equation

In this paper, we introduce a pair of multiplication-like operations, $L_0$ and $L_1$, which derive $k$-regular functions from $(k+1)$-regular functions. The investigation of the inverse problem naturally leads to a deeper study of the 2-Cauchy-Fueter equation. In doing so, we provide a new acyclic resolution for the sheaf of $2$-regular functions $\mathcal{R}^{(2)}$. Furthermore, a complete topological characterization for the solvability of the $2$-Cauchy-Fueter equation is established. Specifically, we prove that the $2$-Cauchy-Fueter equation $$\mathscr{D}^{(2)}f=g$$ is solvable for any $g$ satisfying $\mathscr{D}_1^{(2)}g=0$ on a domain $Ω\subset\mathbb{R}^4$ if and only if $H^3(Ω, \mathbb{R}) = 0$, or equivalently, if and only if every real-valued harmonic function on $Ω$ can be represented as the real part of a quaternionic regular function.

math.CV↗
arXiv2026

Toward Efficient Influence Function: Dropout as a Compression Tool

Assessing the impact the training data on machine learning models is crucial for understanding the behavior of the model, enhancing the transparency, and selecting training data. Influence function provides a theoretical framework for quantifying the effect of training data points on model's performance given a specific test data. However, the computational and memory costs of influence function presents significant challenges, especially for large-scale models, even when using approximation methods, since the gradients involved in computation are as large as the model itself. In this work, we introduce a novel approach that leverages dropout as a gradient compression mechanism to compute the influence function more efficiently. Our method significantly reduces computational and memory overhead, not only during the influence function computation but also in gradient compression process. Through theoretical analysis and empirical validation, we demonstrate that our method could preserves critical components of the data influence and enables its application to modern large-scale models.

cs.LG↗
arXiv2026

NuRisk: A Visual Question Answering Dataset for Agent-Level Risk Assessment in Autonomous Driving

Understanding risk in autonomous driving requires not only perception and prediction, but also high-level reasoning about agent behavior and context. Current Vision Language Model (VLM)-based methods primarily ground agents in static images and provide qualitative judgments, lacking the spatio-temporal reasoning needed to capture how risks evolve over time. To address this gap, we propose NuRisk, a comprehensive Visual Question Answering (VQA) dataset comprising 2.9K scenarios and 1.1M agent-level samples, built on real-world data from nuScenes and Waymo, completed with safety-critical scenarios from the CommonRoad simulator. The dataset provides Bird's-eye view (BEV) based sequential images with quantitative, agent-level risk annotations, enabling spatio-temporal reasoning. We benchmark well-known VLMs across different prompting techniques and find that they fail to perform explicit spatio-temporal reasoning, resulting in a peak accuracy of 33% at high latency. To address these shortcomings, our fine-tuned 7B VLM agent improves accuracy to 41% and reduces latency by 75%, demonstrating explicit spatio-temporal reasoning capabilities that proprietary models lacked. While this represents a significant step forward, the modest accuracy underscores the profound challenge of the task, establishing NuRisk as a critical benchmark for advancing spatio-temporal reasoning in autonomous driving. More information can be found at https://github.com/TUM-AVS/NuRisk.

cs.AI↗
arXiv2026

Target-Bench: Can Video World Models Achieve Mapless Path Planning with Semantic Targets?

While recent video world models can generate highly realistic videos, their ability to perform semantic reasoning and planning remains unclear and unquantified. We introduce Target-Bench, the first benchmark that enables comprehensive evaluation of video world models' semantic reasoning, spatial estimation, and planning capabilities. Target-Bench provides 450 robot-collected scenarios spanning 47 semantic categories, with SLAM-based trajectories serving as motion tendency references. Our benchmark reconstructs motion from generated videos with a metric scale recovery mechanism, enabling the evaluation of planning performance with five complementary metrics that focus on target-approaching capability and directional consistency. Our evaluation result shows that the best off-the-shelf model achieves only a 0.341 overall score, revealing a significant gap between realistic visual generation and semantic reasoning in current video world models. Furthermore, we demonstrate that fine-tuning process on a relatively small real-world robot dataset can significantly improve task-level planning performance.

cs.CV↗
arXiv2026

Weighted $L^2$ theory for the Euclidean Dirac operator in higher dimensions

We study weighted $L^{2}$ solvability for the Euclidean Dirac operator in dimensions $n\ge 3$. We prove that, on the exterior domain $\mathbb{R}^{n}\setminus\overline{B(0,1)}$ with logarithmic weight $φ=n\log|x|$, no higher-dimensional analogue of the two-dimensional Hörmander estimate can be controlled solely by $Δφ$; we then establish weighted solvability for the weights $|x|^{m}$ with $m\neq 0$, for the quadratic weight $x_{1}^{2}$, and for sufficiently small anisotropic perturbations of the Gaussian weight, with sharp constant $1/4$ in the Gaussian case. The obstruction arises because, in dimensions $n\ge 3$, the classical weighted identity is coercive only under a structural relation between $Δφ$ and $|\nablaφ|^{2}$, a condition that excludes the Gaussian weight and many polynomial weights. The method is based on a weighted identity for the conjugated unknown $U:=ue^{-φ/2}$, together with suitable scalar and Clifford-valued multipliers; this identity yields the required coercive estimates and also gives weighted $L^{2}$ solvability for the Poisson equation through the factorization $Δ=-D^{2}$.

math.CV↗
arXiv2026

Near-Field Localization via Reconfigurable Antennas

Reconfigurable antennas (RAs) utilize the electromagnetic (EM) domain to provide dynamic control over antenna radiation patterns, which offers an effective way to enhance power efficiency in wireless links. Unlike conventional arrays with fixed element patterns, RAs enable on-demand beam-pattern synthesis by directly controlling each antenna's EM characteristics. While existing research on RAs has primarily focused on improving spectral efficiency, this paper explores their application for downlink localization. Moreover, the majority of existing works focus on far-field scenarios with little attention on near-field (NF). Motivated by these gaps, we consider a synthesis model in which each antenna generates desired beampatterns from a finite set of EM basis functions. We then formulate a joint optimization problem for the baseband (BB) and EM precoders with the objective of minimizing the user equipment (UE) position error bound (PEB) in NF conditions. Our analytical derivations and extensive simulation results demonstrate that the proposed hybrid precoder design for RAs significantly improves UE positioning accuracy compared to traditional non-reconfigurable arrays.

eess.SP↗
arXiv2026

The Coherence Trap: When MLLM-Crafted Narratives Exploit Manipulated Visual Contexts

The detection and grounding of multimedia manipulation has emerged as a critical challenge in combating AI-generated disinformation. While existing methods have made progress in recent years, we identify two fundamental limitations in current approaches: (1) Underestimation of MLLM-driven deception risk: prevailing techniques primarily address rule-based text manipulations, yet fail to account for sophisticated misinformation synthesized by multimodal large language models (MLLMs) that can dynamically generate semantically coherent, contextually plausible yet deceptive narratives conditioned on manipulated images; (2) Unrealistic misalignment artifacts: currently focused scenarios rely on artificially misaligned content that lacks semantic coherence, rendering them easily detectable. To address these gaps holistically, we propose a new adversarial pipeline that leverages MLLMs to generate high-risk disinformation. Our approach begins with constructing the MLLM-Driven Synthetic Multimodal (MDSM) dataset, where images are first altered using state-of-the-art editing techniques and then paired with MLLM-generated deceptive texts that maintain semantic consistency with the visual manipulations. Building upon this foundation, we present the Artifact-aware Manipulation Diagnosis via MLLM (AMD) framework featuring two key innovations: Artifact Pre-perception Encoding strategy and Manipulation-Oriented Reasoning, to tame MLLMs for the MDSM problem. Comprehensive experiments validate our framework's superior generalization capabilities as a unified architecture for detecting MLLM-powered multimodal deceptions. In cross-domain testing on the MDSM dataset, AMD achieves the best average performance, with 88.18 ACC, 60.25 mAP, and 61.02 mIoU scores.

cs.CV↗
arXiv2026

Holographic Intelligence Surface Assisted Integrated Sensing and Communication

Traditional discrete-array-based systems fail to exploit interactions between closely spaced antennas, resulting in inadequate utilization of the aperture resource. In this paper, we propose a holographic intelligence surface (HIS) assisted integrated sensing and communication (HISAC) system, wherein both the transmitter and receiver are fabricated using a continuous-aperture array. A continuous-discrete transformation of the HIS pattern based on the Fourier transform is proposed, converting the continuous pattern design into a discrete beamforming design. We formulate a joint transmit-receive beamforming optimization problem for the HISAC system, aiming to balance the performance of multi-target sensing while fulfilling the performance requirement of multi-user communication. To solve the non-convex problem with coupled variables, an alternating optimization-based algorithm is proposed to optimize the HISAC transmit-receive beamforming in an alternate manner. Specifically, the transmit beamforming design is solved by decoupling into a series of feasibility-checking sub-problems while the receive beamforming is determined by the Rayleigh quotient-based method. Simulation results demonstrate the superiority of the proposed HISAC system over traditional discrete-array-based ISAC systems, achieving significantly higher sensing performance while guaranteeing predetermined communication performance.

eess.SP↗