arXiv subjects

Xin Zhao

Publications and source records attributed to Xin Zhao.

At least 55 records · Page 3Linked to original sources

arXiv2025

Who Speaks for the Trigger? Dynamic Expert Routing in Backdoored Mixture-of-Experts Transformers

Large language models (LLMs) with Mixture-of-Experts (MoE) architectures achieve impressive performance and efficiency by dynamically routing inputs to specialized subnetworks, known as experts. However, this sparse routing mechanism inherently exhibits task preferences due to expert specialization, introducing a new and underexplored vulnerability to backdoor attacks. In this work, we investigate the feasibility and effectiveness of injecting backdoors into MoE-based LLMs by exploiting their inherent expert routing preferences. We thus propose BadSwitch, a novel backdoor framework that integrates task-coupled dynamic trigger optimization with a sensitivity-guided Top-S expert tracing mechanism. Our approach jointly optimizes trigger embeddings during pretraining while identifying S most sensitive experts, subsequently constraining the Top-K gating mechanism to these targeted experts. Unlike traditional backdoor attacks that rely on superficial data poisoning or model editing, BadSwitch primarily embeds malicious triggers into expert routing paths with strong task affinity, enabling precise and stealthy model manipulation. Through comprehensive evaluations across three prominent MoE architectures (Switch Transformer, QwenMoE, and DeepSeekMoE), we demonstrate that BadSwitch can efficiently hijack pre-trained models with up to 100% success rate (ASR) while maintaining the highest clean accuracy (ACC) among all baselines. Furthermore, BadSwitch exhibits strong resilience against both text-level and model-level defense mechanisms, achieving 94.07% ASR and 87.18% ACC on the AGNews dataset. Our analysis of expert activation patterns reveals fundamental insights into MoE vulnerabilities. We anticipate this work will expose security risks in MoE systems and contribute to advancing AI safety.

cs.CR↗

arXiv2025

Tracing Multilingual Knowledge Acquisition Dynamics in Domain Adaptation: A Case Study of English-Japanese Biomedical Adaptation

Multilingual domain adaptation (ML-DA) is widely used to learn new domain knowledge across languages into large language models (LLMs). Although many methods have been proposed to improve domain adaptation, the mechanisms of multilingual knowledge acquisition, how domain knowledge is learned within a language and transferred across languages, remain underexplored. This gap leads to suboptimal performance, particularly in low-resource settings. This work examines the learning dynamics of LLMs during ML-DA. Because prior ML-DA studies often train and evaluate on datasets with mismatched knowledge coverage, we propose AdaXEval, an adaptive evaluation method that builds multiple-choice QA datasets from the same bilingual domain corpus used for training, thereby directly studying multilingual knowledge acquisition. Through continual training of LLMs with diverse data recipes, we track how LLMs acquire domain facts and pinpoint the mechanism behind the transformation process from domain training data to knowledge. Our experiments on a 13B English-Japanese bilingual LLM reveal that cross-lingual transfer remains challenging despite a high-quality bilingual corpus. The code has been released.

cs.CL↗

arXiv2025

AILoRA: Function-Aware Asymmetric Initialization for Low-Rank Adaptation of Large Language Models

Parameter-efficient finetuning (PEFT) aims to mitigate the substantial computational and memory overhead involved in adapting large-scale pretrained models to diverse downstream tasks. Among numerous PEFT strategies, Low-Rank Adaptation (LoRA) has emerged as one of the most widely adopted approaches due to its robust empirical performance and low implementation complexity. In practical deployment, LoRA is typically applied to the $W^Q$ and $W^V$ projection matrices of self-attention modules, enabling an effective trade-off between model performance and parameter efficiency. While LoRA has achieved considerable empirical success, it still encounters challenges such as suboptimal performance and slow convergence. To address these limitations, we introduce \textbf{AILoRA}, a novel parameter-efficient method that incorporates function-aware asymmetric low-rank priors. Our empirical analysis reveals that the projection matrices $W^Q$ and $W^V$ in the self-attention mechanism exhibit distinct parameter characteristics, stemming from their functional differences. Specifically, $W^Q$ captures task-specific semantic space knowledge essential for attention distributions computation, making its parameters highly sensitive to downstream task variations. In contrast, $W^V$ encodes token-level feature representations that tend to remain stable across tasks and layers. Leveraging these insights, AILoRA performs a function-aware initialization by injecting the principal components of $W^Q$ to retain task-adaptive capacity, and the minor components of $W^V$ to preserve generalizable feature representations. This asymmetric initialization strategy enables LoRA modules to better capture the specialized roles of attention parameters, thereby enhancing both finetuning performance and convergence efficiency.

cs.AI↗

arXiv2025

Break the ID-Language Barrier: An Adaption Framework for LLM-based Sequential Recommendation

The recent breakthrough of large language models (LLMs) in natural language processing has sparked exploration in recommendation systems, however, their limited domain-specific knowledge remains a critical bottleneck. Specifically, LLMs lack key pieces of information crucial for sequential recommendations, such as user behavior patterns. To address this critical gap, we propose IDLE-Adapter, a novel framework that integrates pre-trained ID embeddings, rich in domain-specific knowledge, into LLMs to improve recommendation accuracy. IDLE-Adapter acts as a bridge, transforming sparse user-item interaction data into dense, LLM-compatible representations through a Pre-trained ID Sequential Model, Dimensionality Alignment, Layer-wise Embedding Refinement, and Layer-wise Distribution Alignment. Furthermore, IDLE-Adapter demonstrates remarkable flexibility by seamlessly integrating ID embeddings from diverse ID-based sequential models and LLM architectures. Extensive experiments across various datasets demonstrate the superiority of IDLE-Adapter, achieving over 10\% and 20\% improvements in HitRate@5 and NDCG@5 metrics, respectively, compared to state-of-the-art methods.

cs.IR↗

arXiv2025

Holographic s+p superconductors with axion induced translation symmetry breaking

We construct a holographic model for an s+p superconductor with axion-induced translation symmetry breaking within the framework of gauge/gravity duality, working in the probe limit. The equations of motion are solved numerically to investigate the influence of the parameter $k/T$ on the competition and coexistence between the s-wave and p-wave orders. We find that increasing $k/T$ suppresses the thermodynamic stability of both the single condensate s-wave and p-wave solutions. With the $k-μ$ phase diagram and the condensate curves, we see that the region dominated by the single condensate p-wave phase gradually decreases with the increasing of $k/T$, finally leaving only the single condensate s-wave phase in the large $k/T$ region, which is explained by the grand potential curves showing a slower decreasing of the thermodynamic stability for the s-wave solution than that for the p-wave solution. Furthermore, a larger minimum ratio of the charges $q_p/q_s$ is required to stabilize the s+p coexistent phase as $k/T$ increases, and we determine the precise dependence of this critical ratio on $k/T$. Finally, our study of the optical conductivity reveals that the gap frequency increases with $k/T$. A characteristic kink, associated with the s+p coexistent phase, is identified in the dependence of gap frenquency on $k/T$, which could serve as a potential experimental signature for detecting multi-condensate superconductivity.

hep-th↗

arXiv2025

Structural Equation-VAE: Disentangled Latent Representations for Tabular Data

Learning interpretable latent representations from tabular data remains a challenge in deep generative modeling. We introduce SE-VAE (Structural Equation-Variational Autoencoder), a novel architecture that embeds measurement structure directly into the design of a variational autoencoder. Inspired by structural equation modeling, SE-VAE aligns latent subspaces with known indicator groupings and introduces a global nuisance latent to isolate construct-specific confounding variation. This modular architecture enables disentanglement through design rather than through statistical regularizers alone. We evaluate SE-VAE on a suite of simulated tabular datasets and benchmark its performance against a series of leading baselines using standard disentanglement metrics. SE-VAE consistently outperforms alternatives in factor recovery, interpretability, and robustness to nuisance variation. Ablation results reveal that architectural structure, rather than regularization strength, is the key driver of performance. SE-VAE offers a principled framework for white-box generative modeling in scientific and social domains where latent constructs are theory-driven and measurement validity is essential.

cs.LG↗

arXiv2025

SOI is the Root of All Evil: Quantifying and Breaking Similar Object Interference in Single Object Tracking

In this paper, we present the first systematic investigation and quantification of Similar Object Interference (SOI), a long-overlooked yet critical bottleneck in Single Object Tracking (SOT). Through controlled Online Interference Masking (OIM) experiments, we quantitatively demonstrate that eliminating interference sources leads to substantial performance improvements (AUC gains up to 4.35) across all SOTA trackers, directly validating SOI as a primary constraint for robust tracking and highlighting the feasibility of external cognitive guidance. Building upon these insights, we adopt natural language as a practical form of external guidance, and construct SOIBench-the first semantic cognitive guidance benchmark specifically targeting SOI challenges. It automatically mines SOI frames through multi-tracker collective judgment and introduces a multi-level annotation protocol to generate precise semantic guidance texts. Systematic evaluation on SOIBench reveals a striking finding: existing vision-language tracking (VLT) methods fail to effectively exploit semantic cognitive guidance, achieving only marginal improvements or even performance degradation (AUC changes of -0.26 to +0.71). In contrast, we propose a novel paradigm employing large-scale vision-language models (VLM) as external cognitive engines that can be seamlessly integrated into arbitrary RGB trackers. This approach demonstrates substantial improvements under semantic cognitive guidance (AUC gains up to 0.93), representing a significant advancement over existing VLT methods. We hope SOIBench will serve as a standardized evaluation platform to advance semantic cognitive tracking research and contribute new insights to the tracking research community.

cs.CV↗

arXiv2025

Hide and Seek with LLMs: An Adversarial Game for Sneaky Error Generation and Self-Improving Diagnosis

Large Language Models (LLMs) excel in reasoning and generation across domains, but still struggle with identifying and diagnosing complex errors. This stems mainly from training objectives that prioritize correct answers, limiting exposure to and learning from errors. While recent studies have begun to address this by introducing error signals, most rely on shallow, static errors, restricting improvement in deep diagnostic ability. To overcome this, we propose Hide and Seek Game (HSG), a dynamic adversarial framework for error generation and diagnosis, and evaluate it on mathematical problem-solving. HSG involves two adversarial roles: Sneaky, which "hides" by generating subtle, deceptive reasoning errors, and Diagnosis, which "seeks" to accurately detect them. Through adversarial co-evolution, both error stealth and diagnostic precision are enhanced. Experiments on several math reasoning tasks show that HSG significantly boosts error diagnosis, achieving 16.8\%--31.4\% higher accuracy than baselines like GPT-4o. We also release a challenging dataset of deceptive errors and diagnostic annotations as a benchmark for future research.

cs.AI↗

arXiv2025

Towards Identifying the PL6 Center in SiC: From First-Principles Screening to Hyperfine Validation of Competing Defect Candidates

The PL6 color center in 4H-SiC, known for its excellent ambient-temperature spin and optical properties, has an unresolved microscopic origin. In this first-principles study, we systematically investigate potential structures to clarify its nature. We first rigorously examine the DV-antisite hypothesis (a divacancy paired with a carbon antisite, $\mathrm{C_{Si}}$), analyzing the energetic, electronic, and spin properties of various $V_\mathrm{Si}V_\mathrm{C}+\mathrm{C_{Si}}$ configurations. Two $\mathrm{C_{3v}}$-symmetric $\mathrm{kk+C_{Si}}$ complexes emerge as strong candidates within this framework. Subsequently, a critical comparison of hyperfine interaction signatures is performed between these candidates, the alternative OV model [specifically OV(hh) and OV(kk), an oxygen replacing C together with a Si vacancy], and experimental data. This analysis demonstrates that the OV(hh) structure more accurately reproduces PL6's hyperfine features. Furthermore, re-evaluation of the proposed OV(hk) model for the related PL5 center reveals zero-field splitting parameter $E$ inconsistencies with experimental results, suggesting that PL5 and PL6 may have distinct origins. These findings provide crucial theoretical insights and motivate targeted experimental validation for these quantum defects.

cond-mat.mtrl-sci↗

arXiv2025

A Survey on Vision-Language-Action Models for Autonomous Driving

The rapid progress of multimodal large language models (MLLM) has paved the way for Vision-Language-Action (VLA) paradigms, which integrate visual perception, natural language understanding, and control within a single policy. Researchers in autonomous driving are actively adapting these methods to the vehicle domain. Such models promise autonomous vehicles that can interpret high-level instructions, reason about complex traffic scenes, and make their own decisions. However, the literature remains fragmented and is rapidly expanding. This survey offers the first comprehensive overview of VLA for Autonomous Driving (VLA4AD). We (i) formalize the architectural building blocks shared across recent work, (ii) trace the evolution from early explainer to reasoning-centric VLA models, and (iii) compare over 20 representative models according to VLA's progress in the autonomous driving domain. We also consolidate existing datasets and benchmarks, highlighting protocols that jointly measure driving safety, accuracy, and explanation quality. Finally, we detail open challenges - robustness, real-time efficiency, and formal verification - and outline future directions of VLA4AD. This survey provides a concise yet complete reference for advancing interpretable socially aligned autonomous vehicles. Github repo is available at \href{https://github.com/JohnsonJiang1996/Awesome-VLA4AD}{SicongJiang/Awesome-VLA4AD}.

cs.CV↗

arXiv2025

Progtuning: Progressive Fine-tuning Framework for Transformer-based Language Models

Fine-tuning is a promising technique for leveraging Transformer-based language models in downstream tasks. As model sizes continue to grow, updating all model parameters becomes increasingly costly. Parameter-efficient fine-tuning methods effectively address this issue by selectively updating a small subset of parameters. However, fine-tuning and most existing parameter-efficient fine-tuning methods require updating the same number of parameters as the initial size, ignoring the unequal contribution across Transformer blocks and leading to extremely inefficient allocation of computing resources. In this paper, we propose Progtuning, the novel fine-tuning framework combined with progressive learning for Transformer-based language models. Specifically, Progtuning progressively reduces the number of updated transformer blocks based on the contribution. Remarkably, Progtuning optimizes resource allocation and reduces the number of updated parameters by approximately 25\%, while still maintaining competitive performance. And it also exhibits high adaptability with parameter-efficient fine-tuning methods, demonstrating excellent performance across various adaptation scenarios.

cs.CL↗

arXiv2025

A Literature Review on Simulation in Conversational Recommender Systems

Conversational Recommender Systems (CRSs) have garnered attention as a novel approach to delivering personalized recommendations through multi-turn dialogues. This review developed a taxonomy framework to systematically categorize relevant publications into four groups: dataset construction, algorithm design, system evaluation, and empirical studies, providing a comprehensive analysis of simulation methods in CRSs research. Our analysis reveals that simulation methods play a key role in tackling CRSs' main challenges. For example, LLM-based simulation methods have been used to create conversational recommendation data, enhance CRSs algorithms, and evaluate CRSs. Despite several challenges, such as dataset bias, the limited output flexibility of LLM-based simulations, and the gap between text semantic space and behavioral semantics, persist due to the complexity in Human-Computer Interaction (HCI) of CRSs, simulation methods hold significant potential for advancing CRS research. This review offers a thorough summary of the current research landscape in this domain and identifies promising directions for future inquiry.

cs.HC↗

arXiv2025

Interior structure of the holographic s + p superconductor and chaotic-stable transition near the black hole singularity

In this work, we investigate the interior structure of a holographic multi-band superconductor with the coexistence of s-wave and p-wave order parameters. Especially, we investigate the singularity structure of this multi-band model. Different from the single p-wave case, the alternation rule is jointly determined by parameters involving both s-wave order and p-wave order. In the coexistence region, we derive the Kasner alternation laws from both analytical and numerical methods which fit each other nicely. Furthermore, we find that the occurrence of the s-wave order parameter will lead to a chaotic-stable transition for the near singularity structure which matches the expectation of cosmological billiard approach. This novel transition for the near singularity structure constitutes a holographic counterpart of the secondary condensation in boundary superconducting system, offering a complementary perspective for characterizing the properties of boundary condensed matter systems.

hep-th↗

arXiv2025

Ring-lite: Scalable Reasoning via C3PO-Stabilized Reinforcement Learning for LLMs

We present Ring-lite, a Mixture-of-Experts (MoE)-based large language model optimized via reinforcement learning (RL) to achieve efficient and robust reasoning capabilities. Built upon the publicly available Ling-lite model, a 16.8 billion parameter model with 2.75 billion activated parameters, our approach matches the performance of state-of-the-art (SOTA) small-scale reasoning models on challenging benchmarks (e.g., AIME, LiveCodeBench, GPQA-Diamond) while activating only one-third of the parameters required by comparable models. To accomplish this, we introduce a joint training pipeline integrating distillation with RL, revealing undocumented challenges in MoE RL training. First, we identify optimization instability during RL training, and we propose Constrained Contextual Computation Policy Optimization(C3PO), a novel approach that enhances training stability and improves computational throughput via algorithm-system co-design methodology. Second, we empirically demonstrate that selecting distillation checkpoints based on entropy loss for RL training, rather than validation metrics, yields superior performance-efficiency trade-offs in subsequent RL training. Finally, we develop a two-stage training paradigm to harmonize multi-domain data integration, addressing domain conflicts that arise in training with mixed dataset. We will release the model, dataset, and code.

cs.CL↗

arXiv2025

Diving into a holographic multi-band superconductor

In this work, we investigate the interior structure of a holographic multi-band superconductor model. We focus on the holographic superconductor system with two scalar fields which correspond to two s-wave order parameters in the dual condensed matter system. We discover a new kind of transition rule among Kasner universe near the black hole singularity which is distinct from the holographic single-band superconductor model. This transition rule is the first generalization of Kasner transition behavior to scenarios involving multiple free parameters, which is useful for uncovering the most general interior structures of hairy black holes. Moreover, we find that the Kasner exponents are sensitive to the details of order parameters in boundary system. These different near singularity structures we found show that the black hole interior plays crucial role in characterizing the boundary dual condensed matter systems.

hep-th↗

arXiv2025

Time of Impact Dataset for Continuous Collision Detection and a Scalable Conservative Algorithm

We introduce a large-scale benchmark for broad- and narrow-phase continuous collision detection (CCD) over linearized trajectories with exact time of impacts and use it to evaluate the accuracy, correctness, and efficiency of 13 state-of-the-art CCD algorithms. Our analysis shows that several methods exhibit problems either in efficiency or accuracy. To overcome these limitations, we introduce an algorithm for CCD designed to be scalable on modern parallel architectures and provably correct when implemented using floating point arithmetic. We integrate our algorithm within the Incremental Potential Contact solver [24] and evaluate its impact on various simulation scenarios. Our approach includes a broad-phase CCD to quickly filter out primitives having disjoint bounding boxes and a narrow-phase CCD that establishes whether the remaining primitive pairs indeed collide. Our broad-phase algorithm is efficient and scalable thanks to the experimental observation that sweeping along a coordinate axis performs surprisingly well on modern parallel architectures. For narrow-phase CCD, we re-design the recently proposed interval-based algorithm of Wang et al. [45] to work on massively parallel hardware. To foster the adoption and development of future linear CCD algorithms, and to evaluate their correctness, scalability, and overall performance, we release the dataset with analytic ground truth, the implementation of all the algorithms tested, and our testing framework.

cs.GR↗

arXiv2025

Neuron Empirical Gradient: Discovering and Quantifying Neurons Global Linear Controllability

While feed-forward neurons in pre-trained language models (PLMs) can encode knowledge, past research targeted a small subset of neurons that heavily influence outputs. This leaves the broader role of neuron activations unclear, limiting progress in areas like knowledge editing. We uncover a global linear relationship between neuron activations and outputs using neuron interventions on a knowledge probing dataset. The gradient of this linear relationship, which we call the neuron empirical gradient (NEG), captures how changes in activations affect predictions. To compute NEG efficiently, we propose NeurGrad, enabling large-scale analysis of neuron behavior in PLMs. We also show that NEG effectively captures language skills across diverse prompts through skill neuron probing. Experiments on MCEval8k, a multi-genre multiple-choice knowledge benchmark, support NEG's ability to represent model knowledge. Further analysis highlights the key properties of NEG-based skill representation: efficiency, robustness, flexibility, and interdependency. The code and data are released.

cs.CL↗

arXiv2025

RMoA: Optimizing Mixture-of-Agents through Diversity Maximization and Residual Compensation

Although multi-agent systems based on large language models show strong capabilities on multiple tasks, they are still limited by high computational overhead, information loss, and robustness. Inspired by ResNet's residual learning, we propose Residual Mixture-of-Agents (RMoA), integrating residual connections to optimize efficiency and reliability. To maximize information utilization from model responses while minimizing computational costs, we innovatively design an embedding-based diversity selection mechanism that greedily selects responses via vector similarity. Furthermore, to mitigate iterative information degradation, we introduce a Residual Extraction Agent to preserve cross-layer incremental information by capturing inter-layer response differences, coupled with a Residual Aggregation Agent for hierarchical information integration. Additionally, we propose an adaptive termination mechanism that dynamically halts processing based on residual convergence, further improving inference efficiency. RMoA achieves state-of-the-art performance on the benchmarks of across alignment, mathematical reasoning, code generation, and multitasking understanding, while significantly reducing computational overhead. Code is available at https://github.com/mindhunter01/RMoA.

cs.AI↗