arXiv subjects

Ze Wang

Publications and source records attributed to Ze Wang.

At least 55 records · Page 3Linked to original sources

arXiv2025

Unleashing Hour-Scale Video Training for Long Video-Language Understanding

Recent long-form video-language understanding benchmarks have driven progress in video large multimodal models (Video-LMMs). However, the scarcity of well-annotated long videos has left the training of hour-long Video-LMMs underexplored. To close this gap, we present VideoMarathon, a large-scale hour-long video instruction-following dataset. This dataset includes around 9,700 hours of long videos sourced from diverse domains, ranging from 3 to 60 minutes per video. Specifically, it contains 3.3M high-quality QA pairs, spanning six fundamental topics: temporality, spatiality, object, action, scene, and event. Compared to existing video instruction datasets, VideoMarathon significantly extends training video durations up to 1 hour, and supports 22 diverse tasks requiring both short- and long-term video comprehension. Building on VideoMarathon, we propose Hour-LLaVA, a powerful and efficient Video-LMM for hour-scale video-language modeling. It enables hour-long video training and inference at 1-FPS sampling by leveraging a memory augmentation module, which adaptively integrates question-relevant and spatiotemporally informative semantics from the cached full video context. In our experiments, Hour-LLaVA achieves the best performance on multiple representative long video-language benchmarks, demonstrating the high quality of the VideoMarathon dataset and the superiority of the Hour-LLaVA model.

cs.CV↗

arXiv2025

Integrated soliton microcombs beyond the turnkey limit

Soliton microcombs generated in optical microresonators are accelerating the transition of optical frequency combs from laboratory instruments to industrial platforms. Self injection locking (SIL) enables direct driving of soliton microcombs by integrated lasers, providing turnkey initiation and improved coherence, but it also pins the pump close to resonance, limiting both spectral span and tuning flexibility. Here we theoretically and experimentally demonstrate that introducing a thermally tunable auxiliary microresonator extends the bandwidth of SIL soliton microcombs. By engineering hybridization of the pumped resonance, we achieve deterministic access to single soliton states and then push operation into a far detuned regime inaccessible to direct initiation. The resulting combs reach a near 200 nm span at a 25 GHz repetition rate, while preserving the SIL-enabled noise suppression throughout. Moreover, the added degree of freedom afforded by the coupled resonator architecture enables orthogonal control of the comb's repetition rate and center frequency. These advances expand the spectral reach and controllability of integrated soliton microcombs for information processing and precision metrology.

physics.optics↗

arXiv2025

Instella: Fully Open Language Models with Stellar Performance

Large language models (LLMs) have demonstrated remarkable performance across a wide range of tasks, yet the majority of high-performing models remain closed-source or partially open, limiting transparency and reproducibility. In this work, we introduce Instella, a family of fully open three billion parameter language models trained entirely on openly available data and codebase. Powered by AMD Instinct MI300X GPUs, Instella is developed through large-scale pre-training, general-purpose instruction tuning, and alignment with human preferences. Despite using substantially fewer pre-training tokens than many contemporaries, Instella achieves state-of-the-art results among fully open models and is competitive with leading open-weight models of comparable size. We further release two specialized variants: Instella-Long, capable of handling context lengths up to 128K tokens, and Instella-Math, a reasoning-focused model enhanced through supervised fine-tuning and reinforcement learning on mathematical tasks. Together, these contributions establish Instella as a transparent, performant, and versatile alternative for the community, advancing the goal of open and reproducible language modeling research.

cs.CL↗

arXiv2025

Learning from Online Videos at Inference Time for Computer-Use Agents

Computer-use agents can operate computers and automate laborious tasks, but despite recent rapid progress, they still lag behind human users, especially when tasks require domain-specific procedural knowledge about particular applications, platforms, and multi-step workflows. Humans can bridge this gap by watching video tutorials: we search, skim, and selectively imitate short segments that match our current subgoal. In this paper, we study how to enable computer-use agents to learn from online videos at inference time effectively. We propose a framework that retrieves and filters tutorial videos, converts them into structured demonstration trajectories, and dynamically selects trajectories as in-context guidance during execution. Particularly, using a VLM, we infer UI actions, segment videos into short subsequences of actions, and assign each subsequence a textual objective. At inference time, a two-stage selection mechanism dynamically chooses a single trajectory to add in context at each step, focusing the agent on the most helpful local guidance for its next decision. Experiments on two widely used benchmarks show that our framework consistently outperforms strong base agents and variants that use only textual tutorials or transcripts. Analyses highlight the importance of trajectory segmentation and selection, action filtering, and visual information, suggesting that abundant online videos can be systematically distilled into actionable guidance that improves computer-use agents at inference time. Our code is available at https://github.com/UCSB-NLP-Chang/video_demo.

cs.CV↗

arXiv2025

Adversarial Distilled Retrieval-Augmented Guarding Model for Online Malicious Intent Detection

With the deployment of Large Language Models (LLMs) in interactive applications, online malicious intent detection has become increasingly critical. However, existing approaches fall short of handling diverse and complex user queries in real time. To address these challenges, we introduce ADRAG (Adversarial Distilled Retrieval-Augmented Guard), a two-stage framework for robust and efficient online malicious intent detection. In the training stage, a high-capacity teacher model is trained on adversarially perturbed, retrieval-augmented inputs to learn robust decision boundaries over diverse and complex user queries. In the inference stage, a distillation scheduler transfers the teacher's knowledge into a compact student model, with a continually updated knowledge base collected online. At deployment, the compact student model leverages top-K similar safety exemplars retrieved from the online-updated knowledge base to enable both online and real-time malicious query detection. Evaluations across ten safety benchmarks demonstrate that ADRAG, with a 149M-parameter model, achieves 98.5% of WildGuard-7B's performance, surpasses GPT-4 by 3.3% and Llama-Guard-3-8B by 9.5% on out-of-distribution detection, while simultaneously delivering up to 5.6x lower latency at 300 queries per second (QPS) in real-time applications.

cs.CR↗

arXiv2025

KeyVID: Keyframe-Aware Video Diffusion for Audio-Synchronized Visual Animation

Generating video from various conditions, such as text, image, and audio, enables both spatial and temporal control, leading to high-quality generation results. Videos with dramatic motions often require a higher frame rate to ensure smooth motion. Currently, most audio-to-visual animation models use uniformly sampled frames from video clips. However, these uniformly sampled frames fail to capture significant key moments in dramatic motions at low frame rates and require significantly more memory when increasing the number of frames directly. In this paper, we propose KeyVID, a keyframe-aware audio-to-visual animation framework that significantly improves the generation quality for key moments in audio signals while maintaining computation efficiency. Given an image and an audio input, we first localize keyframe time steps from the audio. Then, we use a keyframe generator to generate the corresponding visual keyframes. Finally, we generate all intermediate frames using the motion interpolator. Through extensive experiments, we demonstrate that KeyVID significantly improves audio-video synchronization and video quality across multiple datasets, particularly for highly dynamic motions. The code is released in https://github.com/XingruiWang/KeyVID.

cs.CV↗

arXiv2025

Latent Visual Reasoning

Multimodal Large Language Models (MLLMs) have achieved notable gains in various tasks by incorporating Chain-of-Thought (CoT) reasoning in language spaces. Recent work extends this direction by leveraging external tools for visual editing, thereby enhancing the visual signal along the reasoning trajectories. Nevertheless, these approaches remain fundamentally constrained: reasoning is still confined to the language space, with visual information treated as static preconditions. We introduce Latent Visual Reasoning (LVR), a new paradigm that enables autoregressive reasoning directly in the visual embedding space. A visual encoder first projects images into visual tokens within a joint semantic space shared with the language model. The language model is then trained to generate latent states that reconstruct key visual tokens critical for answering the query, constituting the process of latent visual reasoning. By interleaving LVR with standard text generation, our model achieves substantial gains on perception-intensive visual question answering tasks. In addition, we adapt the GRPO algorithm to conduct reinforcement learning on latent reasoning, further balancing LVR and textual generation. We show that LVR substantially improves fine-grained visual understanding and perception, achieving 71.67% on MMVP compared to 66.67% with Qwen2.5-VL. Code base and model weights will be released later.

cs.CV↗

arXiv2025

Electrically-pumped soliton microcombs on thin-film lithium niobate

Thin-film lithium niobate (TFLN) has enabled efficient on-chip electro-optic modulation and frequency conversion for information processing and precision measurement. Extending these capabilities with optical frequency combs unlocks massively parallel operations and coherent optical-to-microwave transduction, which are achievable in TFLN microresonators via Kerr microcombs. However, fully integrated Kerr microcombs directly driven by semiconductor lasers remain elusive, which has delayed integration of these technologies. Here we demonstrate electrically pumped TFLN Kerr microcombs without optical amplification. With optimized laser-to-chip coupling and optical quality factors, we generate soliton microcombs at a 200 GHz repetition frequency with an optical span of 180 nm using only 25 mW of pump power. Moreover, self-injection locking enables turnkey initiation and substantially narrows the laser linewidth. Our work provides integrated comb sources for TFLN-based communicational, computational, and metrological applications.

physics.optics↗

arXiv2025

ImageDoctor: Diagnosing Text-to-Image Generation via Grounded Image Reasoning

The rapid advancement of text-to-image (T2I) models has increased the need for reliable human preference modeling, a demand further amplified by recent progress in reinforcement learning for preference alignment. However, existing approaches typically quantify the quality of a generated image using a single scalar, limiting their ability to provide comprehensive and interpretable feedback on image quality. To address this, we introduce ImageDoctor, a unified multi-aspect T2I model evaluation framework that assesses image quality across four complementary dimensions: plausibility, semantic alignment, aesthetics, and overall quality. ImageDoctor also provides pixel-level flaw indicators in the form of heatmaps, which highlight misaligned or implausible regions, and can be used as a dense reward for T2I model preference alignment. Inspired by the diagnostic process, we improve the detail sensitivity and reasoning capability of ImageDoctor by introducing a "look-think-predict" paradigm, where the model first localizes potential flaws, then generates reasoning, and finally concludes the evaluation with quantitative scores. Built on top of a vision-language model and trained through a combination of supervised fine-tuning and reinforcement learning, ImageDoctor demonstrates strong alignment with human preference across multiple datasets, establishing its effectiveness as an evaluation metric. Furthermore, when used as a reward model for preference tuning, ImageDoctor significantly improves generation quality -- achieving an improvement of 10% over scalar-based reward models.

cs.CV↗

arXiv2025

APRIL: Active Partial Rollouts in Reinforcement Learning to Tame Long-tail Generation

Reinforcement learning (RL) has become a cornerstone in advancing large-scale pre-trained language models (LLMs). Successive generations, including GPT-o series, DeepSeek-R1, Kimi-K1.5, Grok 4, and GLM-4.5, have relied on large-scale RL training to enhance reasoning and coding capabilities. To meet the community's growing RL needs, numerous RL frameworks have been proposed. However, RL training remains computationally expensive, with rollout generation accounting for more than 90% of total runtime. In addition, its efficiency is often constrained by the long-tail distribution of rollout response lengths, where a few lengthy responses stall entire batches, leaving GPUs idle and underutilized. As model and rollout sizes continue to grow, this bottleneck increasingly limits scalability. To address this challenge, we propose Active Partial Rollouts in Reinforcement Learning (APRIL), which mitigates long-tail inefficiency. In the rollout phase, APRIL over-provisions rollout requests, terminates once the target number of responses is reached, and recycles incomplete responses for continuation in future steps. This strategy ensures that no rollouts are discarded while substantially reducing GPU idle time. Experiments show that APRIL improves rollout throughput by 22.5% on average (at most 44%) across commonly used RL algorithms (GRPO, DAPO, GSPO), accelerates convergence, and achieves 2.1% on average(at most 8%) higher final accuracy across tasks. Moreover, APRIL is both framework and hardware agnostic, already integrated into the slime RL framework, and deployable on NVIDIA and AMD GPUs alike. Taken together, this work unifies system-level and algorithmic considerations in proposing APRIL, with the aim of advancing RL training efficiency and inspiring further optimizations in RL systems. Our codebase is available at https://github.com/RLsys-Foundation/APRIL

cs.LG↗

arXiv2025

Knowledge Collapse in LLMs: When Fluency Survives but Facts Fail under Recursive Synthetic Training

Large language models increasingly rely on synthetic data due to human-written content scarcity, yet recursive training on model-generated outputs leads to model collapse, a degenerative process threatening factual reliability. We define knowledge collapse as a distinct three-stage phenomenon where factual accuracy deteriorates while surface fluency persists, creating "confidently wrong" outputs that pose critical risks in accuracy-dependent domains. Through controlled experiments with recursive synthetic training, we demonstrate that collapse trajectory and timing depend critically on instruction format, distinguishing instruction-following collapse from traditional model collapse through its conditional, prompt-dependent nature. We propose domain-specific synthetic training as a targeted mitigation strategy that achieves substantial improvements in collapse resistance while maintaining computational efficiency. Our evaluation framework combines model-centric indicators with task-centric metrics to detect distinct degradation phases, enabling reproducible assessment of epistemic deterioration across different language models. These findings provide both theoretical insights into collapse dynamics and practical guidance for sustainable AI training in knowledge-intensive applications where accuracy is paramount.

cs.CL↗

arXiv2025

The Sound of Risk: A Multimodal Physics-Informed Acoustic Model for Forecasting Market Volatility and Enhancing Market Interpretability

Information asymmetry in financial markets, often amplified by strategically crafted corporate narratives, undermines the effectiveness of conventional textual analysis. We propose a novel multimodal framework for financial risk assessment that integrates textual sentiment with paralinguistic cues derived from executive vocal tract dynamics in earnings calls. Central to this framework is the Physics-Informed Acoustic Model (PIAM), which applies nonlinear acoustics to robustly extract emotional signatures from raw teleconference sound subject to distortions such as signal clipping. Both acoustic and textual emotional states are projected onto an interpretable three-dimensional Affective State Label (ASL) space-Tension, Stability, and Arousal. Using a dataset of 1,795 earnings calls (approximately 1,800 hours), we construct features capturing dynamic shifts in executive affect between scripted presentation and spontaneous Q&A exchanges. Our key finding reveals a pronounced divergence in predictive capacity: while multimodal features do not forecast directional stock returns, they explain up to 43.8% of the out-of-sample variance in 30-day realized volatility. Importantly, volatility predictions are strongly driven by emotional dynamics during executive transitions from scripted to spontaneous speech, particularly reduced textual stability and heightened acoustic instability from CFOs, and significant arousal variability from CEOs. An ablation study confirms that our multimodal approach substantially outperforms a financials-only baseline, underscoring the complementary contributions of acoustic and textual modalities. By decoding latent markers of uncertainty from verifiable biometric signals, our methodology provides investors and regulators a powerful tool for enhancing market interpretability and identifying hidden corporate uncertainty.

cs.LG↗

arXiv2025

Enhanced Superconductivity and Mixed-dimensional Behaviour in Infinite-layer Samarium Nickelate Thin Films

Rare-earth infinite-layer nickelates represent an emerging class of unconventional superconductors, with materials synthesis largely limited to early lanthanide compounds. Here, we report the synthesis and characterization of phase-pure superconducting samarium-based infinite-layer nickelate thin films, including the first demonstration of Sm$_{1-x}$Sr$_x$NiO$_2$, along with co-doped variants incorporating europium and calcium. These films, grown on LSAT (001) substrates, exhibit coherent lattice structures up to $\sim$ 9 nm thickness with minimal stacking faults. The co-doped compounds achieve a record-small $c$-axis parameter of 3.26 Å and display remarkable superconducting transition temperatures up to 32.5 K. These results establish a clear correlation between decreasing $c$-axis parameter and increasing critical temperature across different rare-earth systems. In addition, angle-dependent magnetoresistance investigations reveal the existence of a hybrid mixture of 2D and 3D superconductivity in this novel system with enhanced coupling between the rare-earth 5d and Ni 3d orbitals, confirmed by resonant inelastic X-ray scattering experiments. As the concentration of Eu increases, the system exhibits a clear tendency towards 3D superconductivity. Furthermore, we observe distinctive negative magnetoresistance in the europium-containing samples. These findings advocate clear materials design principles for higher transition temperatures and exotic physics in infinite-layer nickelate superconductors through structural engineering of the rare-earth site.

cond-mat.supr-con↗

arXiv2025

P/D-Device: Disaggregated Large Language Model between Cloud and Devices

Serving disaggregated large language models has been widely adopted in industrial practice for enhanced performance. However, too many tokens generated in decoding phase, i.e., occupying the resources for a long time, essentially hamper the cloud from achieving a higher throughput. Meanwhile, due to limited on-device resources, the time to first token (TTFT), i.e., the latency of prefill phase, increases dramatically with the growth on prompt length. In order to concur with such a bottleneck on resources, i.e., long occupation in cloud and limited on-device computing capacity, we propose to separate large language model between cloud and devices. That is, the cloud helps a portion of the content for each device, only in its prefill phase. Specifically, after receiving the first token from the cloud, decoupling with its own prefill, the device responds to the user immediately for a lower TTFT. Then, the following tokens from cloud are presented via a speed controller for smoothed TPOT (the time per output token), until the device catches up with the progress. On-device prefill is then amortized using received tokens while the resource usage in cloud is controlled. Moreover, during cloud prefill, the prompt can be refined, using those intermediate data already generated, to further speed up on-device inference. We implement such a scheme P/D-Device, and confirm its superiority over other alternatives. We further propose an algorithm to decide the best settings. Real-trace experiments show that TTFT decreases at least 60%, maximum TPOT is about tens of milliseconds, and cloud throughput increases by up to 15x.

cs.DC↗

arXiv2025

Power-efficient ultra-broadband soliton microcombs in resonantly-coupled microresonators

The drive to miniaturize optical frequency combs for practical deployment has spotlighted microresonator solitons as a promising chip-scale candidate. However, these soliton microcombs could be very power-hungry when their span increases, especially with fine comb spacings. As a result, realizing an octave-spanning comb at microwave repetition rates for direct optical-microwave linkage is considered not possible for photonic integration due to the high power requirements. Here, we introduce the concept of resonant-coupling to soliton microcombs to reduce pump consumption significantly. Compared to conventional waveguide-coupled designs, we demonstrate (i) a threefold increase in spectral span for high-power combs and (ii) up to a tenfold reduction in repetition frequency for octave-spanning operation. This configuration is compatible with laser integration and yields reliable, turnkey soliton generation. By eliminating the long-standing pump-power bottleneck, microcombs will soon become readily available for portable optical clocks, massively parallel data links, and field-deployable spectrometers.

physics.optics↗

arXiv2025

SF-TIM: A Simple Framework for Enhancing Quadrupedal Robot Jumping Agility by Combining Terrain Imagination and Measurement

Dynamic jumping on high platforms and over gaps differentiates legged robots from wheeled counterparts. Dynamic locomotion on abrupt surfaces, as opposed to walking on rough terrains, demands the integration of proprioceptive and exteroceptive perception to enable explosive movements. In this paper, we propose SF-TIM (Simple Framework combining Terrain Imagination and Measurement), a single-policy method that enhances quadrupedal robot jumping agility, while preserving their fundamental blind walking capabilities. In addition, we introduce a terrain-guided reward design specifically to assist quadrupedal robots in high jumping, improving their performance in this task. To narrow the simulation-to-reality gap in quadrupedal robot learning, we introduce a stable and high-speed elevation map generation framework, enabling zero-shot simulation-to-reality transfer of locomotion ability. Our algorithm has been deployed and validated on both the small-/large-size quadrupedal robots, demonstrating its effectiveness in real-world applications: the robot has successfully traversed various high platforms and gaps, showing the robustness of our proposed approach. A demo video has been made available at https://flysoaryun.github.io/SF-TIM.

cs.RO↗

arXiv2025

MPF: Aligning and Debiasing Language Models post Deployment via Multi Perspective Fusion

Multiperspective Fusion (MPF) is a novel posttraining alignment framework for large language models (LLMs) developed in response to the growing need for easy bias mitigation. Built on top of the SAGED pipeline, an automated system for constructing bias benchmarks and extracting interpretable baseline distributions, MPF leverages multiperspective generations to expose and align biases in LLM outputs with nuanced, humanlike baselines. By decomposing baseline, such as sentiment distributions from HR professionals, into interpretable perspective components, MPF guides generation through sampling and balancing of responses, weighted by the probabilities obtained in the decomposition. Empirically, we demonstrate its ability to align LLM sentiment distributions with both counterfactual baselines (absolute equality) and the HR baseline (biased for Top Univeristy), resulting in small KL divergence, reduction of calibration error and generalization to unseen questions. This shows that MPF offers a scalable and interpretable method for alignment and bias mitigation, compatible with deployed LLMs and requiring no extensive prompt engineering or finetuning.

cs.CL↗

arXiv2025

DocShaDiffusion: Diffusion Model in Latent Space for Document Image Shadow Removal

Document shadow removal is a crucial task in the field of document image enhancement. However, existing methods tend to remove shadows with constant color background and ignore color shadows. In this paper, we first design a diffusion model in latent space for document image shadow removal, called DocShaDiffusion. It translates shadow images from pixel space to latent space, enabling the model to more easily capture essential features. To address the issue of color shadows, we design a shadow soft-mask generation module (SSGM). It is able to produce accurate shadow mask and add noise into shadow regions specially. Guided by the shadow mask, a shadow mask-aware guided diffusion module (SMGDM) is proposed to remove shadows from document images by supervising the diffusion and denoising process. We also propose a shadow-robust perceptual feature loss to preserve details and structures in document images. Moreover, we develop a large-scale synthetic document color shadow removal dataset (SDCSRD). It simulates the distribution of realistic color shadows and provides powerful supports for the training of models. Experiments on three public datasets validate the proposed method's superiority over state-of-the-art. Our code and dataset will be publicly available.

cs.CV↗