arXiv subjects

Xingjun Wang

Publications and source records attributed to Xingjun Wang.

At least 19 recordsLinked to original sources

arXiv2026

LENS: In-Context Search via Latent Evidence Exploration over Dynamic Raw Documents

LLM agents increasingly answer questions over dynamic raw-document collections, where files may change before preprocessing, and relevant evidence (spans, sections, pages, or tables) is query-dependent. Existing retrieval-augmented approaches pre-materialize evidence via fixed chunking, embeddings, or persistent indexes: effective for lookup, yet costly, stale-prone, and committed to a granularity before the query is known. We formulate in-context search as Budgeted Evidence Localization over a latent evidence space induced by dynamic raw documents and propose LENS (Latent Evidence Exploration and Search), an index-free framework. Instead of pre-materializing the evidence space, LENS maintains a query-conditioned belief over candidate units, iteratively selecting candidates via complementary lexical, local, and exploratory proposal policies, updating the belief via an LLM relevance oracle, and narrowing toward high-posterior regions under a controllable budget. Evidence is consolidated into compact, source-grounded regions of interest and compressed into self-organizing knowledge clusters reused across related queries. On a controlled 500-question evaluation with matched corpus snapshots, LENS reaches 62.4% exact match and 84.8% evidence recall vs. 65.2% exact match but 50.4% evidence recall for a ReAct-style baseline. Across scales, LENS gives the strongest supporting-fact localization and answer grounding. On a fixed 150-question fullwiki subset over the raw Wikipedia dump with zero indexing, LENS and ReAct are nearly tied in official answer quality (43.3% vs. 42.7% EM), with LENS grounding more answers in retrieved evidence (84.0% vs. 70.7%). A no-retrieval Closed-Book reference highlights the contribution of model memory. LENS is query-ready after corpus changes, needs no preprocessing or persistent index, and preserves source-grounded evidence localization throughout.

cs.CL↗

arXiv2026

Wan-Animate-2: Pushing the Application Boundaries of Character Animation

Character image animation remains a foundational yet challenging task in computer vision. Existing approaches can be broadly categorized into three paradigms: methods based on explicit motion representations suffer from extraction errors and identity drift; methods based on implicit motion features lose fine-grained dynamics through compression; and in-context learning approaches avoid intermediate representations but incur prohibitive computational costs. Furthermore, all current systems are designed for offline synthesis, unable to meet the real-time requirements of interactive applications such as digital avatars and live-streaming hosts. To address these limitations, we present Wan-Animate-2, an end-to-end character animation framework that directly consumes the driving video within a redesigned Diffusion Transformer. Our architecture achieves superior motion fidelity and identity preservation by eliminating intermediate motion extractors entirely. We further introduce text driven viewpoint control that decouples the output camera perspective from the driving video--a capability rarely supported by prior character animation methods that rely on explicit motion representations. Beyond generation quality, we present Wan-Animate-2-Lite, an efficient variant that reduces inference latency to real-time thresholds through a three-stage training paradigm: teacher forcing pretraining with error buffer mechanism, and Self-Forcing distillation with chunk-wise backpropagation. This enables streaming character animation for interactive applications, opening new deployment scenarios that were previously infeasible. Qualitative evaluations and user studies demonstrate that Wan-Animate-2 achieves high-fidelity animation results across diverse characters and motion patterns. To foster further research and community development, we will release the Wan-Animate-2-Base model weights to the public.

cs.CV↗

arXiv2026

Which Tokens Matter? Adaptive Token Selection for RLVR with the Relative Surprisal Index

Reinforcement learning (RL) has become a powerful tool for propelling Large Language Models (LLMs) beyond imitation-based training towards more robust reasoning capabilities. Among existing approaches, RL with Verifiable Rewards (RLVR) has emerged as a pivotal paradigm for advancing LLM reasoning. Despite its empirical success, recent studies have offered different insights. One line of inquiry advocates prioritizing high-entropy token positions during training, while another perspective cautions against allowing low-probability tokens to dominate gradient updates. Notably, although high-entropy tokens are usually correlated with low probability, both paradigms empirically yield substantial performance gains. In this work, we argue that evaluating sampled-token probability or entropy in isolation is insufficient to capture the policy optimization dynamics. To resolve this tension, we introduce the Relative Surprisal Index (RSI), a principled, information-theoretic metric that naturally couples the token's entropy with the probability of the selected token. We show that, under mild conditions, RSI is related to the local ratio between the first-order variations of the logit-gradient norm and predictive entropy under a selected-logit perturbation. Building on RSI, we propose RSI Selection (RSI-S), an entropy-adaptive token filtering method that retains tokens within a stable RSI interval. RSI-S successfully reconciles previous contradictory paradigms and filters out both redundant low-surprisal tokens and unstable high-surprisal tail tokens. Empirical evaluations show that RSI-S achieves higher avg@32 accuracy across different model scales (Qwen2.5-1.5B, 3B, and 7B) on AIME and AMC benchmarks: RSI-S improves avg@32 accuracy by 2--3 percentage points over GRPO. Overall, RSI offers a promising perspective for RLVR improvement.

cs.AI↗

arXiv2026

Integrated lithium niobate microwave photonics: Driving next-generation wireless technologies

Integrated microwave photonics (MWP) offers a powerful paradigm for handling high-speed microwave signals within chip-scale optical systems. It provides a cost-effective solution to address bandwidth, tunability, and loss bottlenecks of electronics-based radio frequency (RF) systems. The recently emerged thin-film lithium niobate (TFLN) photonic platform, with its exceptional electro-optic (EO) properties, low loss, and scalability, has shown promise to reshape the MWP landscape. Here, we discuss the performance implications of state-of-the-art TFLN photonic devices for MWP applications and offer insights into the emerging trends for next-generation wireless networks. In particular, the unparalleled EO bandwidth enables direct optical generation, processing, and reception of millimeter-wave or even terahertz (THz) signals, significantly expanding the operation frequency range of MWP systems. The low drive voltages and linearity of TFLN modulators lead to an unprecedented operation regime of radio-over-fiber (RoF) systems, featuring net gain, low noise figure and large dynamic range, simultaneously. The availability of a versatile device toolkit, combined with low optical loss and scalability, further supports the transition from traditional tabletop MWP systems to chip-scale solutions, with advanced functionalities, compact footprint, and enhanced system robustness. As the TFLN industrial ecosystem rapidly matures, TFLN-based MWP technology has the potential to deliver transformative solutions to future 6G integrated sensing and communication networks.

physics.optics↗

arXiv2026

GroupGuard: A Framework for Modeling and Defending Collusive Attacks in Multi-Agent Systems

While large language model-based agents demonstrate great potential in collaborative tasks, their interactivity also introduces security vulnerabilities. In this paper, we propose and model group collusive attacks, a highly destructive threat in which multiple agents coordinate via sociological strategies to mislead the system. To address this challenge, we introduce GroupGuard, a training-free defense framework that employs a multi-layered defense strategy, including continuous graph-based monitoring, active honeypot inducement, and structural pruning, to identify and isolate collusive agents. Experimental results across five datasets and four topologies demonstrate that group collusive attacks increase the attack success rate by up to 15\% compared to individual attacks. GroupGuard consistently achieves high detection accuracy (up to 88\%) and effectively restores collaborative performance, providing a robust solution for securing multi-agent systems.

cs.AI↗

arXiv2025

Self-sustained microcomb lasing in an integrated hybrid oscillator

Microcavity optical frequency combs (microcombs) are compact, coherent light sources whose chip-scale integrability is poised to drive advances in metrology, communications, and sensing. Among available microcomb generation methods, hybrid cavities uniquely co-locate gain and Kerr dynamics, where the lasing mode directly resonates in the nonlinear microcavity, simultaneously enabling self-sustained and highly efficient microcomb generation. However, their implementation is often limited by partial integration or the need for external injection, which complicates operation architecture, raises power and hampers system miniaturization. In this work, we present a fully integrated hybrid cavity for self-sustained microcomb generation, relying solely on the co-oscillation of lasing and Kerr nonlinearity without external driving. The system collapses the pump laser, nonlinear resonator and feedback loops into a minimalist on-chip two-element cavity, consisting of a high-Q microresonator with engineered intracavity reflection and a reflective semiconductor optical amplifier (RSOA). The scheme delivers self-starting operation and stable performance without active feedback. The generated coherent microcomb achieves intrinsic linewidths below 1 kHz and integrated linewidths around 100 kHz, with self-sustained operation exceeding 24 hours. This ultra-compact architecture provides a practical path toward scalable, coherent multi-wavelength sources for integrated photonic systems.

physics.optics↗

arXiv2025

HarmRLVR: Weaponizing Verifiable Rewards for Harmful LLM Alignment

Recent advancements in Reinforcement Learning with Verifiable Rewards (RLVR) have gained significant attention due to their objective and verifiable reward signals, demonstrating strong performance in reasoning and code generation tasks. However, the potential safety risks associated with RLVR remain underexplored. This paper presents HarmRLVR, the first systematic investigation into the alignment reversibility risk of RLVR. We show that safety alignment can be rapidly reversed using GRPO with merely 64 harmful prompts without responses, causing models to readily comply with harmful instructions. Across five models from Llama, Qwen, and DeepSeek, we empirically demonstrate that RLVR-based attacks elevate the average harmfulness score to 4.94 with an attack success rate of 96.01\%, significantly outperforming harmful fine-tuning while preserving general capabilities. Our findings reveal that RLVR can be efficiently exploited for harmful alignment, posing serious threats to open-source model safety. Please see our code at https://github.com/lyxx2535/HarmRLVR.

cs.CR↗

arXiv2025

Integrated photonic ultrawideband real-time spectrum sensing for 6G wireless networks

The sixth generation (6G) wireless networks require dynamic spectrum management to optimize the utilization of scarce spectral resources and support emerging integrated sensing and communication (ISAC) applications. This necessitates real-time spectrum sensing (RT-SS) capability with ultrawide measurement range, compact size, and low latency. Conventional electronic RT-SS solutions face critical challenges in operating across the millimeter-wave and sub-terahertz bands, which are essential spectra for 6G wireless. While photonic RT-SS has the potential to surpass this limitation, the current implementations feature limited bandwidths below 50 GHz and mostly rely on bulky dispersive fibers with high latency. Here, we address these challenges by developing an integrated photonic RT-SS system capable of ultrabroadband measurement from microwave to sub-terahertz bands, covering the full spectrum for 6G wireless. The photonic RT-SS circuit integrates a broadband electro-optic (EO) modulator for unknown signal loading, an EO tunable microring filter bank for high-speed parallel frequency-to-time mapping, as well as an EO comb for precise channel frequency referencing, all realized on a single thin-film lithium niobate chip. We achieve an unprecedented spectral measurement range beyond 120 GHz at a low latency of less than 100 ns. To validate the effectiveness of our photonic RT-SS system in future 6G scenes, we further propose a heuristic spectro-temporal resource allocation algorithm and conduct a proof-of-concept ISAC demonstration, where a radar adaptively access RT-SS-informed spectrally underutilized regions for high-quality target sensing under dynamic communication interferences. Our work presents a compact and cost-effective solution for efficient spectrum sharing and dynamic management in future 6G ISAC networks.

physics.optics↗

arXiv2025

Ultrabroadband Integrated Photonics Empowering Full-Spectrum Adaptive Wireless Communications

The forthcoming sixth-generation (6G) and beyond (XG) wireless networks are poised to operate across an expansive frequency range from microwave, millimeter-wave to terahertz bands to support ubiquitous connectivity in diverse application scenarios. This necessitates a one-size-fits-all hardware solution that can be adaptively reconfigured within this wide spectrum to support full-band coverage and dynamic spectrum management. However, existing electrical or photonic-assisted wireless communication solutions see significant challenges in meeting this demand due to the limited bandwidths of individual devices and the intrinsically rigid nature of their system architectures. Here, we demonstrate adaptive wireless communications over an unprecedented frequency range spanning over 100 GHz, driven by a universal thin-film lithium niobate (TFLN) photonic wireless engine. Leveraging the strong Pockels effect and excellent scalability of the TFLN platform, we achieve monolithic integration of essential functional elements, including baseband modulation, broadband wireless-photonic conversion, and reconfigurable carrier/local signal generation. Powered by broadband tunable optoelectronic oscillators, our signal sources operate across a record-wide frequency range from 0.5 GHz to 115 GHz with high frequency stability and consistent coherence. Based on the broadband and reconfigurable integrated photonic solution, we realize, for the first time, full-link wireless communication across 9 consecutive bands, achieving record lane speeds of up to 100 Gbps. The real-time reconfigurability further enables adaptive frequency allocation, a crucial capability to ensure enhanced reliability in complex spectrum environments. Our proposed system marks a significant step towards future full-spectrum and omni-scenario wireless networks.

physics.optics↗

arXiv2025

GDGS: 3D Gaussian Splatting Via Geometry-Guided Initialization And Dynamic Density Control

We propose a method to enhance 3D Gaussian Splatting (3DGS)~\cite{Kerbl2023}, addressing challenges in initialization, optimization, and density control. Gaussian Splatting is an alternative for rendering realistic images while supporting real-time performance, and it has gained popularity due to its explicit 3D Gaussian representation. However, 3DGS heavily depends on accurate initialization and faces difficulties in optimizing unstructured Gaussian distributions into ordered surfaces, with limited adaptive density control mechanism proposed so far. Our first key contribution is a geometry-guided initialization to predict Gaussian parameters, ensuring precise placement and faster convergence. We then introduce a surface-aligned optimization strategy to refine Gaussian placement, improving geometric accuracy and aligning with the surface normals of the scene. Finally, we present a dynamic adaptive density control mechanism that adjusts Gaussian density based on regional complexity, for visual fidelity. These innovations enable our method to achieve high-fidelity real-time rendering and significant improvements in visual quality, even in complex scenes. Our method demonstrates comparable or superior results to state-of-the-art methods, rendering high-fidelity images in real time.

cs.CV↗

arXiv2025

Nexus-Gen: Unified Image Understanding, Generation, and Editing via Prefilled Autoregression in Shared Embedding Space

Unified multimodal generative models aim to integrate image understanding and generation abilities, offering significant advantages in harnessing multimodal corpora, particularly interleaved text-image data. However, existing unified models exhibit limitations in image synthesis quality, autoregressive error accumulation, and image editing capability. In this work, we propose Nexus-Gen, a novel architecture that unifies image understanding, generation, and editing tasks in a shared image embedding space. This shared space serves as a bridge for the autoregressive and diffusion models, which seamlessly integrates their complementary strengths in cross-modal modeling. To mitigate the severe error accumulation during autoregressive embedding prediction, we propose a novel prefilled autoregression strategy that aligns training-inference dynamics by prefilling input sequences with learnable embeddings. After multi-stage and multi-task training on our constructed large-scale dataset with 26.3 million samples, Nexus-Gen achieves state-of-the-art performance on the evaluation benchmarks spanning image understanding, generation and editing tasks. All models, datasets, and source codes are released in https://github.com/modelscope/Nexus-Gen to facilitate further advancements across the field.

cs.CV↗

arXiv2025

Breaking the Modality Barrier: Universal Embedding Learning with Multimodal LLMs

The Contrastive Language-Image Pre-training (CLIP) framework has become a widely used approach for multimodal representation learning, particularly in image-text retrieval and clustering. However, its efficacy is constrained by three key limitations: (1) text token truncation, (2) isolated image-text encoding, and (3) deficient compositionality due to bag-of-words behavior. While recent Multimodal Large Language Models (MLLMs) have demonstrated significant advances in generalized vision-language understanding, their potential for learning transferable multimodal representations remains underexplored.In this work, we present UniME (Universal Multimodal Embedding), a novel two-stage framework that leverages MLLMs to learn discriminative representations for diverse downstream tasks. In the first stage, we perform textual discriminative knowledge distillation from a powerful LLM-based teacher model to enhance the embedding capability of the MLLM\'s language component. In the second stage, we introduce hard negative enhanced instruction tuning to further advance discriminative representation learning. Specifically, we initially mitigate false negative contamination and then sample multiple hard negatives per instance within each batch, forcing the model to focus on challenging samples. This approach not only improves discriminative power but also enhances instruction-following ability in downstream tasks. We conduct extensive experiments on the MMEB benchmark and multiple retrieval tasks, including short and long caption retrieval and compositional retrieval. Results demonstrate that UniME achieves consistent performance improvement across all tasks, exhibiting superior discriminative and compositional capabilities.

cs.CV↗

arXiv2025

EliGen: Entity-Level Controlled Image Generation with Regional Attention

Recent advancements in diffusion models have significantly advanced text-to-image generation, yet global text prompts alone remain insufficient for achieving fine-grained control over individual entities within an image. To address this limitation, we present EliGen, a novel framework for Entity-level controlled image Generation. Firstly, we put forward regional attention, a mechanism for diffusion transformers that requires no additional parameters, seamlessly integrating entity prompts and arbitrary-shaped spatial masks. By contributing a high-quality dataset with fine-grained spatial and semantic entity-level annotations, we train EliGen to achieve robust and accurate entity-level manipulation, surpassing existing methods in both spatial precision and image quality. Additionally, we propose an inpainting fusion pipeline, extending its capabilities to multi-entity image inpainting tasks. We further demonstrate its flexibility by integrating it with other open-source models such as IP-Adapter, In-Context LoRA and MLLM, unlocking new creative possibilities. The source code, model, and dataset are published at https://github.com/modelscope/DiffSynth-Studio.git.

cs.CV↗

arXiv2024

Minimum Tuning to Unlock Long Output from LLMs with High Quality Data as the Key

As large language models rapidly evolve to support longer context, there is a notable disparity in their capability to generate output at greater lengths. Recent study suggests that the primary cause for this imbalance may arise from the lack of data with long-output during alignment training. In light of this observation, attempts are made to re-align foundation models with data that fills the gap, which result in models capable of generating lengthy output when instructed. In this paper, we explore the impact of data-quality in tuning a model for long output, and the possibility of doing so from the starting points of human-aligned (instruct or chat) models. With careful data curation, we show that it possible to achieve similar performance improvement in our tuned models, with only a small fraction of training data instances and compute. In addition, we assess the generalizability of such approaches by applying our tuning-recipes to several models. our findings suggest that, while capacities for generating long output vary across different models out-of-the-box, our approach to tune them with high-quality data using lite compute, consistently yields notable improvement across all models we experimented on. We have made public our curated dataset for tuning long-writing capability, the implementations of model tuning and evaluation, as well as the fine-tuned models, all of which can be openly-accessed.

cs.CL↗

arXiv2024

Integrated adaptive coherent LiDAR for 4D bionic vision

Light detection and ranging (LiDAR) is a ubiquitous tool to provide precise spatial awareness in various perception environments. A bionic LiDAR that can mimic human-like vision by adaptively gazing at selected regions of interest within a broad field of view is crucial to achieve high-resolution imaging in an energy-saving and cost-effective manner. However, current LiDARs based on stacking fixed-wavelength laser arrays and inertial scanning have not been able to achieve the desired dynamic focusing patterns and agile scalability simultaneously. Moreover, the ability to synchronously acquire multi-dimensional physical parameters, including distance, direction, Doppler, and color, through seamless fusion between multiple sensors, still remains elusive in LiDAR. Here, we overcome these limitations and demonstrate a bio-inspired frequency-modulated continuous wave (FMCW) LiDAR system with dynamic and scalable gazing capability. Our chip-scale LiDAR system is built using hybrid integrated photonic solutions, where a frequency-chirped external cavity laser provides broad spectral tunability, while on-chip electro-optic combs with elastic channel spacing allow customizable imaging granularity. Using the dynamic zoom-in capability and the coherent FMCW scheme, we achieve a state-of-the-art resolution of 0.012 degrees, providing up to 15 times the resolution of conventional 3D LiDAR sensors, with 115 equivalent scanning lines and 4D parallel imaging. We further demonstrate cooperative sensing between our adaptive coherent LiDAR and a camera to enable high-resolution color-enhanced machine vision.

physics.optics↗

arXiv2024

SWIFT:A Scalable lightWeight Infrastructure for Fine-Tuning

Recent development in Large Language Models (LLMs) and Multi-modal Large Language Models (MLLMs) have leverage Attention-based Transformer architectures and achieved superior performance and generalization capabilities. They have since covered extensive areas of traditional learning tasks. For instance, text-based tasks such as text-classification and sequence-labeling, as well as multi-modal tasks like Visual Question Answering (VQA) and Optical Character Recognition (OCR), which were previously addressed using different models, can now be tackled based on one foundation model. Consequently, the training and lightweight fine-tuning of LLMs and MLLMs, especially those based on Transformer architecture, has become particularly important. In recognition of these overwhelming needs, we develop SWIFT, a customizable one-stop infrastructure for large models. With support of over $300+$ LLMs and $50+$ MLLMs, SWIFT stands as the open-source framework that provide the most comprehensive support for fine-tuning large models. In particular, it is the first training framework that provides systematic support for MLLMs. In addition to the core functionalities of fine-tuning, SWIFT also integrates post-training processes such as inference, evaluation, and model quantization, to facilitate fast adoptions of large models in various application scenarios. With a systematic integration of various training techniques, SWIFT offers helpful utilities such as benchmark comparisons among different training techniques for large models. For fine-tuning models specialized in agent framework, we show that notable improvements on the ToolBench leader-board can be achieved by training with customized dataset on SWIFT, with an increase of 5.2%-21.8% in the Act.EM metric over various baseline models, a reduction in hallucination by 1.6%-14.1%, and an average performance improvement of 8%-17%.

cs.CL↗

arXiv2023

High-coherence parallelization in integrated photonics

Coherent optics has profoundly impacted diverse applications ranging from communications, LiDAR to quantum computations. However, building coherent systems in integrated photonics previously came at great expense in hardware integration and energy efficiency: the lack of a power-efficient way to generate highly coherent light necessitates bulky lasers and amplifiers, while frequency and phase recovery schemes require huge digital signal processing resources. In this work, we demonstrate a high-coherence parallelization strategy that facilitates advanced integrated coherent systems at a minimum price. Using a self-injection locked microcomb to injection lock a distributed feedback laser array, we boost the microcomb power by a record high gain of up to 60 dB on chip with no degradation in coherence. This strategy enables tens of highly coherent channels with an intrinsic linewidth down to the 10 Hz level and power of more than 20 dBm. The overall electrical to optical wall-plug efficiency reaches 19%, comparable with that of the state-of-the-art semiconductor lasers. Driven by this parallel source, we demonstrate a silicon photonic communication link with an unprecedented data rate beyond 60 Tbit/s. Importantly, the high coherence we achieve reduces the coherent-related DSP consumption by 99.999% compared with the traditional III-V laser pump scheme. This work paves a way to realizing scalable, high-performance coherent integrated photonic systems, potentially benefiting numerous applications.

physics.optics↗

arXiv2023

FaceChain: A Playground for Human-centric Artificial Intelligence Generated Content

Recent advancement in personalized image generation have unveiled the intriguing capability of pre-trained text-to-image models on learning identity information from a collection of portrait images. However, existing solutions are vulnerable in producing truthful details, and usually suffer from several defects such as (i) The generated face exhibit its own unique characteristics, \ie facial shape and facial feature positioning may not resemble key characteristics of the input, and (ii) The synthesized face may contain warped, blurred or corrupted regions. In this paper, we present FaceChain, a personalized portrait generation framework that combines a series of customized image-generation model and a rich set of face-related perceptual understanding models (\eg, face detection, deep face embedding extraction, and facial attribute recognition), to tackle aforementioned challenges and to generate truthful personalized portraits, with only a handful of portrait images as input. Concretely, we inject several SOTA face models into the generation procedure, achieving a more efficient label-tagging, data-processing, and model post-processing compared to previous solutions, such as DreamBooth ~\cite{ruiz2023dreambooth} , InstantBooth ~\cite{shi2023instantbooth} , or other LoRA-only approaches ~\cite{hu2021lora} . Besides, based on FaceChain, we further develop several applications to build a broader playground for better showing its value, including virtual try on and 2D talking head. We hope it can grow to serve the burgeoning needs from the communities. Note that this is an ongoing work that will be consistently refined and improved upon. FaceChain is open-sourced under Apache-2.0 license at \url{https://github.com/modelscope/facechain}.

cs.CV↗