arXiv subjects

Xiang Shen

Publications and source records attributed to Xiang Shen.

17 recordsLinked to original sources

arXiv2026

EvoHarness-RL: Learning Self-Evolving Runtime Harness for Long-Horizon LLM Agents

Long-horizon LLM agents increasingly rely on external execution support to maintain state, track progress, invoke tools, verify outcomes, and reuse experience across interactions. However, effective harness use raises two coupled challenges: state formation from noisy interaction traces and runtime control over external-state access. Existing agents usually handle both through prompts, heuristics, or domain-specific conventions, leaving the external workspace and its usage policy manually engineered. To address this, we study the problem of harness policy learning, where agents learn harness policies offline and deploy them to construct and update external harness state online during runtime task execution. We introduce EvoHarness-RL, which exposes Belief, Progress, and Experience (BPE) as policy-facing harness state. Supervised harness fine-tuning teaches the base agent the harness action space and how to construct useful external state, while cost-aware GRPO explores coordination policies to selectively read, update, and consolidate that state during long-horizon interaction. Instantiated on ALFWorld with a Qwen3-8B LLM, EvoHarness-RL reaches 96.9% success and reveals two key dynamics: harness annealing, where training internalizes recurring harness-use patterns into the model policy and shifts the agent from frequent harness calls toward selective external-state access, and harness evolution, where progress updates and experience consolidation refine the harness into a compact, task-adaptive state substrate. These results suggest that long-horizon agents benefit from trainable policies for constructing and coordinating with external harness workspaces, beyond simply adding stronger tools or larger memories.

cs.LG↗

arXiv2026

Agentic Recommender System with Hierarchical Belief-State Memory

Memory-augmented LLM agents have advanced personalized recommendation, yet existing approaches universally adopt flat memory representations that conflate ephemeral signals with stable preferences, and none provides a complete lifecycle governing how memory should evolve. We propose MARS (Memory-Augmented Agentic Recommender System), a framework that treats recommendation as a partially observable problem and maintains a structured belief state that progressively abstracts noisy behavioral observations into a compact estimate of user preferences. MARS organizes this belief state into three tiers: event memory buffers raw signals, preference memory maintains fine-grained mutable chunks with explicit strength and evidence tracking, and profile memory distills all preferences into a coherent natural language narrative. A complete lifecycle of six operations -- extraction, reinforcement, weakening, consolidation, forgetting, and resynthesis -- is adaptively scheduled by an LLM-based planner rather than fixed-interval heuristics. Experiments on four InstructRec benchmark domains show that MARS achieves state-of-the-art performance with average improvements of 26.4% in HR@1 and 10.3% in NDCG@10 over the strongest baselines with further gains from agentic scheduling in evolving settings.

cs.CL↗

arXiv2026

Switchable high-Q light absorbers based on phase-change resonant metasurfaces

In this paper, we propose a switchable high-Q light absorber based on a reconfigurable metasurface enabled by a lowloss phase-change material (PCM). By leveraging the coupling between guided-mode resonance and Fabry-Perot modes, mediated by the phase-transition dynamics of the embedded PCM, the resonance Q factor can be actively tuned. This allows the system to switch from a perfect dark state, governed by the physics of bound states in the continuum, to a critically coupled resonance with a finite Q factor. Consequently, the metasurface exhibits perfect absorption in the amorphous state and a reflection-dominated response in the crystalline state. The proposed metasurface holds significant potential for diverse nanophotonic applications, including photodetection and thermal emission control.

physics.optics↗

arXiv2025

Reasoning-Enhanced Domain-Adaptive Pretraining of Multimodal Large Language Models for Short Video Content Governance

Short video platforms are evolving rapidly, making the identification of inappropriate content increasingly critical. Existing approaches typically train separate and small classification models for each type of issue, which requires extensive human-labeled data and lacks cross-issue generalization. We propose a reasoning-enhanced multimodal large language model (MLLM) pretraining paradigm for unified inappropriate content detection. To address the distribution gap between short video content and the original pretraining data of MLLMs, as well as the complex issue definitions, we introduce three targeted pretraining tasks: (1) \textit{Caption}, to enhance the MLLM's perception of video details; (2) \textit{Visual Question Answering (VQA)}, to deepen the MLLM's understanding of issue definitions and annotation guidelines; (3) \textit{Chain-of-Thought (CoT)}, to enhance the MLLM's reasoning capability. Experimental results show that our pretraining approach significantly improves the MLLM's performance in both zero-shot and supervised fine-tuning (SFT) settings. In addition, our pretrained model demonstrates strong generalization capabilities to emergent, previously unseen issues.

cs.CV↗

arXiv2025

SFGNet: Semantic and Frequency Guided Network for Camouflaged Object Detection

Camouflaged object detection (COD) aims to segment objects that blend into their surroundings. However, most existing studies overlook the semantic differences among textual prompts of different targets as well as fine-grained frequency features. In this work, we propose a novel Semantic and Frequency Guided Network (SFGNet), which incorporates semantic prompts and frequency-domain features to capture camouflaged objects and improve boundary perception. We further design Multi-Band Fourier Module(MBFM) to enhance the ability of the network in handling complex backgrounds and blurred boundaries. In addition, we design an Interactive Structure Enhancement Block (ISEB) to ensure structural integrity and boundary details in the predictions. Extensive experiments conducted on three COD benchmark datasets demonstrate that our method significantly outperforms state-of-the-art approaches. The core code of the model is available at the following link: https://github.com/winter794444/SFGNetICASSP2026.

cs.CV↗

arXiv2025

Compact polarization-independent non-volatile optical switches

Compact, non-volatile optical switches on silicon platforms are essential for reconfigurable photonics, but the strong anisotropy of silicon waveguides leads to polarization-dependent performance. In this paper, we propose a polarization-independent, non-volatile optical switch utilizing low-loss phase change material (PCM) Sb2S3. By incorporating Sb2S3 into a multimode slot waveguide, multimode interference can be efficiently tuned for both TE and TM polarizations, owing to enhanced light-PCM interaction. Polarization-independent switching is achieved through the optimal design of the multimode slot waveguide region. The proposed non-volatile switch demonstrates a crosstalk (CT) < -21.9 dB and insertion loss (IL) < 0.12 dB at 1550 nm with a multimode section length of 9.67 {\mu}m, which may find promising applications in reconfigurable photonic circuits for on-chip optical signal processing.

physics.optics↗

arXiv2025

Filter-And-Refine: A MLLM Based Cascade System for Industrial-Scale Video Content Moderation

Effective content moderation is essential for video platforms to safeguard user experience and uphold community standards. While traditional video classification models effectively handle well-defined moderation tasks, they struggle with complicated scenarios such as implicit harmful content and contextual ambiguity. Multimodal large language models (MLLMs) offer a promising solution to these limitations with their superior cross-modal reasoning and contextual understanding. However, two key challenges hinder their industrial adoption. First, the high computational cost of MLLMs makes full-scale deployment impractical. Second, adapting generative models for discriminative classification remains an open research problem. In this paper, we first introduce an efficient method to transform a generative MLLM into a multimodal classifier using minimal discriminative training data. To enable industry-scale deployment, we then propose a router-ranking cascade system that integrates MLLMs with a lightweight router model. Offline experiments demonstrate that our MLLM-based approach improves F1 score by 66.50% over traditional classifiers while requiring only 2% of the fine-tuning data. Online evaluations show that our system increases automatic content moderation volume by 41%, while the cascading deployment reduces computational cost to only 1.5% of direct full-scale deployment.

cs.LG↗

arXiv2025

Embedding-based Retrieval in Multimodal Content Moderation

Video understanding plays a fundamental role for content moderation on short video platforms, enabling the detection of inappropriate content. While classification remains the dominant approach for content moderation, it often struggles in scenarios requiring rapid and cost-efficient responses, such as trend adaptation and urgent escalations. To address this issue, we introduce an Embedding-Based Retrieval (EBR) method designed to complement traditional classification approaches. We first leverage a Supervised Contrastive Learning (SCL) framework to train a suite of foundation embedding models, including both single-modal and multi-modal architectures. Our models demonstrate superior performance over established contrastive learning methods such as CLIP and MoCo. Building on these embedding models, we design and implement the embedding-based retrieval system that integrates embedding generation and video retrieval to enable efficient and effective trend handling. Comprehensive offline experiments on 25 diverse emerging trends show that EBR improves ROC-AUC from 0.85 to 0.99 and PR-AUC from 0.35 to 0.95. Further online experiments reveal that EBR increases action rates by 10.32% and reduces operational costs by over 80%, while also enhancing interpretability and flexibility compared to classification-based solutions.

cs.IR↗

arXiv2025

Audio-Enhanced Vision-Language Modeling with Latent Space Broadening for High Quality Data Expansion

Transformer-based multimodal models are widely used in industrial-scale recommendation, search, and advertising systems for content understanding and relevance ranking. Enhancing labeled training data quality and cross-modal fusion significantly improves model performance, influencing key metrics such as quality view rates and ad revenue. High-quality annotations are crucial for advancing content modeling, yet traditional statistical-based active learning (AL) methods face limitations: they struggle to detect overconfident misclassifications and are less effective in distinguishing semantically similar items in deep neural networks. Additionally, audio information plays an increasing role, especially in short-video platforms, yet most pre-trained multimodal architectures primarily focus on text and images. While training from scratch across all three modalities is possible, it sacrifices the benefits of leveraging existing pre-trained visual-language (VL) and audio models. To address these challenges, we propose kNN-based Latent Space Broadening (LSB) to enhance AL efficiency and Vision-Language Modeling with Audio Enhancement (VLMAE), a mid-fusion approach integrating audio into VL models. This system deployed in production systems, leading to significant business gains.

cs.MM↗

arXiv2024

IPS: In-Prompt Process Supervision for Short Video Content Moderation

Multimodal large language models (MLLMs) are effective at capturing the semantics of short video content; however, they often fail to attend to the policy-specific details required for reliable content moderation. To address this limitation, we introduce IPS, a novel framework that integrates In-prompt Process Supervision into MLLMs by introducing sequential reasoning over ancillary questions during fine-tuning. IPS consistently outperforms baseline MLLMs on public and proprietary benchmarks. Moreover, replacing human-annotated ancillary labels with MLLM-generated ones results in only marginal performance degradation, demonstrating robustness to noisy supervision and strong scalability with model-generated annotations. These findings establish IPS as a scalable and effective solution for complex multimodal classification in large-scale industrial settings.

cs.CL↗

arXiv2024

Degenerate merging BICs in resonant metasurfaces

Resonant metasurfaces driven by bound states in the continuum (BIC) offer an intriguing approach to engineer high-Q resonances. Merging multiple BICs in the momentum space could further enhance the Q-factor as well as its robustness to fabrication imperfections. Here, we report doubly-degenerate guided mode resonances (GMR) in a resonant metasurface, whose radiation losses could be totally suppressed due to merging BICs. We show that the GMRs and their associated accidental BICs can be evolved into degenerate merging BICs by parametric tuning of the metasurface. Significantly, these two GMRs share the same critical parameter (i.e. lattice constants or thickness) that the merging BICs occur. Interestingly, thanks to the degenerate property of two GMRs, a larger (smaller) period will split one of merging BICs into eight accidental BICs at off-{\Gamma} point, but annihilate the other. Such exotic phenomenon can be well explained from the interaction of GMRs and background Fabry-Perot resonances. Our result provides new strategies to engineering high-Q resonances in resonant metasurfaces for light-matter interaction.

physics.optics↗

arXiv2024

CPFD: Confidence-aware Privileged Feature Distillation for Short Video Classification

Dense features, customized for different business scenarios, are essential in short video classification. However, their complexity, specific adaptation requirements, and high computational costs make them resource-intensive and less accessible during online inference. Consequently, these dense features are categorized as `Privileged Dense Features'.Meanwhile, end-to-end multi-modal models have shown promising results in numerous computer vision tasks. In industrial applications, prioritizing end-to-end multi-modal features, can enhance efficiency but often leads to the loss of valuable information from historical privileged dense features. To integrate both features while maintaining efficiency and manageable resource costs, we present Confidence-aware Privileged Feature Distillation (CPFD), which empowers features of an end-to-end multi-modal model by adaptively distilling privileged features during training. Unlike existing privileged feature distillation (PFD) methods, which apply uniform weights to all instances during distillation, potentially causing unstable performance across different business scenarios and a notable performance gap between teacher model (Dense Feature enhanced multimodal-model DF-X-VLM) and student model (multimodal-model only X-VLM), our CPFD leverages confidence scores derived from the teacher model to adaptively mitigate the performance variance with the student model. We conducted extensive offline experiments on five diverse tasks demonstrating that CPFD improves the video classification F1 score by 6.76% compared with end-to-end multimodal-model (X-VLM) and by 2.31% with vanilla PFD on-average. And it reduces the performance gap by 84.6% and achieves results comparable to teacher model DF-X-VLM. The effectiveness of CPFD is further substantiated by online experiments, and our framework has been deployed in production systems for over a dozen models.

cs.LG↗

arXiv2022

Q-factor mediated quasi-BIC resonances coupling in asymmetric dimer lattices

Resonance coupling in the regime of bound states in the continuum (BICs) provides an efficient method for engineering nanostructure's optical response with various lineshape while maintaining an ultra-narrow linewidth feature, where the quality factor of resonances plays a crucial role. Independent manipulation of the Q factors of BIC resonances enables full control of interaction behavior as well as both near- and far-field light engineering. In this paper, we harness reflection symmetry (RS) and translational symmetry (TS) protected BIC resonances supported in an asymmetric dimer lattice and investigate Q-factor-mediated resonance coupling behavior under controlled TS and RS perturbations. We focus on in-plane electrical dipole BIC (EDi-BIC) and magnetic dipole BIC (MD-BIC) which are protected by RS, and out-of-plane electrical dipole BIC (EDo-BIC) protected by TS. The coupling between EDi-BIC and EDo-BIC exhibits a resonance crossing behavior where the transmission spectrum at the crossing could be tuned flexibly, showing an electromagnetically induced transparency lineshape or satisfying the lattice Kerker condition with pure phase modulation capability depending on TS and RS perturbed Q factors. While the coupling between MD-BIC and EDo-BIC shows an avoided resonance crossing behavior, where the strongly coupled resonances would lead to the formation of a Friedrich-Wintgen BICs whose spectral position could also be shifted by tuning the Q factors. Our results suggest an intriguing platform to explore BIC resonance interactions with independent Q factor manipulation capability for realizing multi-functional meta-devices.

physics.optics↗

arXiv2022

Dark modes governed by translational-symmetry-protected bound states in the continuum in symmetric dimer lattices

Creating nonradiating dark modes is key to achieving high-Q resonance in dielectric open cavities. The concept of photonic bound states in the continuum (BIC) offers an efficient method to suppress radiative loss through symmetry engineering. Structural reflection symmetry (RS) has been widely utilized to construct BICs in asymmetric metasurfaces. In this paper, we show that the radiation channel of translational-symmetry (TS) protected BIC in 1D symmetric dimer lattice could be unlocked by dimer spacing perturbation. A semi-analytical coupled mode analysis reveals that the total radiation suppression of the TS-BIC is due to the elimination of the first Fourier harmonic component in the lattice parameters. TS-BIC mechanism could also be applied in a 2D symmetric dimer lattice, and BICs protected by TS are robust to RS breaking, and vice versa, providing a promising way to independently control the quality factor of two interacting BIC resonances. Our results suggest a new degree of freedom to engineer BICs as well as their interactions in dimer lattices tailored by different symmetries, and could provide new insight for realizing practical applications requiring high-Q resonances.

physics.optics↗

arXiv2022

Plug-Play Plasmonic Metafibers for Ultrafast Fiber Lasers

Metafibers expand the functionalities of conventional optical fibers to unprecedented nanoscale light manipulations by integrating metasurfaces on the fiber tips, becoming an emerging light-coupling platform for both nanoscience and fiber optics communities. Mostly exploring the isolated bare fibers, current metafibers remain as proof-of-concept demonstrations due to a lack of standard interfaces with the universal fiber networks. Here, we develop new methodologies to fabricate well-defined plasmonic metasurfaces directly on the end facets of commercial single mode fiber jumpers using standard planar technologies and provide a first demonstration of their practical applications in the nonlinear optics regime. Featuring plug-play connections with fiber circuitry and arbitrary metasurfaces landscapes, the metafibers with tunable plasmonic resonances are implemented into fiber laser cavities, yielding all-fiber sub-picosecond (minimum 513 fs) soliton mode locked lasers at optical wavelengths of 1.5 micrometer and 2 micrometer, demonstrating their unusual polarimetric nonlinear transfer functions and superior saturation absorption responses. Novel insights into the physical mechanisms behind the saturable absorption of plasmonic metasurfaces are provided. The nanofabrication process flow is compatible with existing cleanroom technologies, offering metafibers an avenue to be a regular member of functionalized fiber components. The work paves the way towards next generation of ultrafast fiber lasers, optical frequency combs, optical neural networks and ultracompact "all-in-fibers" optical systems for sensing, imaging, communications, and many others.

physics.optics↗

arXiv2021

Tracking Fast Neural Adaptation by Globally Adaptive Point Process Estimation for Brain-Machine Interface

Brain-machine interfaces (BMIs) help the disabled restore body functions by translating neural activity into digital commands to control external devices. Neural adaptation, where the brain signals change in response to external stimuli or movements, plays an important role in BMIs. When subjects purely use neural activity to brain-control a prosthesis, some neurons will actively explore a new tuning property to accomplish the movement task. The prediction of this neural tuning property can help subjects adapt more efficiently to brain control and maintain good decoding performance. Existing prediction methods track the slow change of the tuning property in the manual control, which is not suitable for the fast neural adaptation in brain control. In order to identify the active neurons in brain control and track their tuning property changes, we propose a globally adaptive point process method (GaPP) to estimate the neural modulation state from spike trains, decompose the states into the hyper preferred direction and reconstruct the kinematics in a dual-model framework. We implement the method on real data from rats performing a two-lever discrimination task under manual control and brain control. The results show our method successfully predicts the neural modulation state and identifies the neurons that become active in brain control. Compared to existing methods, ours tracks the fast changes of the hyper preferred direction from manual control to brain control more accurately and efficiently and reconstructs the kinematics better and faster.

eess.SP↗

arXiv1992

Quantum Dilaton Gravity in the Light-cone Gauge

Recently, models of two-dimensional dilaton gravity have been shown to admit classical black-hole solutions that exhibit Hawking radiation at the semi-classical level. These classical and semi-classical analyses have been performed in conformal gauge. We show in this paper that a similar analysis in the light--cone gauge leads to the same results. Moreover, quantization of matter fields in light--cone gauge can be naturally extended to include quantizing the metric field {\it à la} KPZ. We argue that this may provide a new framework to address many issues associated to black-hole physics.

hep-th↗