Searcharxiv⌕ Search

arXiv subjects

Yuhao Zhu

Publications and source records attributed to Yuhao Zhu.

At least 19 recordsLinked to original sources

Reference Forcing: Long Video Generation with Region-Guided Drift Correction

Long video generation should keep the subject and background consistent while supporting rich, natural motion. However, autoregressive models are often distilled with distribution matching distillation (DMD) over short windows, which does not train them to keep this consistency over long ranges. As a result, the subject and background can drift over time. Methods that manage the key-value (KV) cache retain or retrieve history to reduce drift. Yet this history may itself have drifted. We propose Reference Forcing, which uses the initial frames as an appearance reference. We first build a teacher that uses attention masks to force each region of a new frame to attend to the same region in the initial frames. The teacher finds these regions with off-the-shelf segmentation models. We then distill this behavior into a separate attention path of the student, which needs no external models at inference. We further introduce value correction, which corrects drifted values in the KV cache toward the matching values of drift-free initial frames. Experiments show that Reference Forcing achieves state-of-the-art subject and background consistency among long-video methods on different base models, and its advantage grows with video length. Ablations show both components are needed to gain consistency without losing motion. More visualizations and code are available at https://zhaosheng-thu.github.io/reference-force/

cs.CV↗

PrivMark: Pixel-Space Steering against Privacy Leakage in Vision-Language Models

Vision-language models (VLMs) routinely disclose soft-biometric attributes such as age, gender, and race in their text outputs, exposing demographic information from uploaded images to any downstream API user. Existing privacy defenses do not fit this setting: face de-identification leaves distributed soft-biometric cues intact, classifier-targeted adversarial perturbations do not transfer to open-ended generation, and stronger VLM-specific methods require retraining or runtime activation access unavailable to downstream APIs. We propose PrivMark, a framework that suppresses biometric leakage at the image level. By analyzing VLM activations, we identify two weakly aligned steering vector directions in hidden-state space, each tied to a distinct leakage behavior --- one corresponding to incidental attribute disclosure under open-ended descriptions, the other to compliance with direct biometric probes instead of refusal --- and encode both directions into a single bounded pixel perturbation, so the protected image alone enforces privacy with no invasive access to model internals at deployment. To support evaluation, we assemble BioleakBench, $1{,}600$ VQA pairs covering biometric privacy attributes under both free-form description and direct-probe protocols. Across three white-box source models, PrivMark achieves up to $85.9\%$ biometric suppression while retaining $75\%$ of clean OK-VQA utility, up to $1.4\times$ the privacy-utility tradeoff of the strongest pixel-only baseline. The same one-shot perturbations transfer consistently to seven unseen VLMs including closed-source APIs.

cs.CV↗

GLoD-LVSM: Learning Geometry-Grounded Continuous-LoD Scene Representations for Large View Synthesis Models

Large view synthesis models (LVSMs) reconstruct scenes from sparse views and render novel views quickly, but they incur unnecessary storage and rendering costs due to token redundancy. We present GLoD-LVSM, a compact, geometry-grounded scene representation for LVSMs with continuous levels of detail. First, instead of learning a fixed set of scene anchor initializations, GLoD-LVSM learns a continuous field that initializes a flexible number of anchors to match the desired budget. A condenser then aggregates scene information into these anchors, with both the field and condenser trained end-to-end. Second, the condenser infers reference-token and anchor geometry to guide compression through geometry-grounded attention, improving rendering quality and enabling frustum culling during decoding. Compared with existing LVSM scene-compression baselines, GLoD-LVSM achieves the best trade-off between rendering quality and scene token count. Compared with the uncompressed base model, GLoD-LVSM reduces the scene-token count by 33.4% and accelerates rendering by $1.23\times$ with no quality loss. Under a relaxed quality-loss budget, the reduction increases to 85.5%, with $3.00\times$ rendering speedup. Further analysis reveals that view-synthesis supervision yields emergent scene depth and self-organized anchors that align with surfaces and adapt their density to local texture complexity. Beyond rendering, we further show that our scene representation also supports scene-semantic understanding, highlighting its potential for broader downstream tasks.

cs.CV↗

HoOPOE: A Holography-Optimized Perceptual ObjectivE function

Holographic devices must reconstruct an entire light field given a finite number of spatial light modulator (SLM) pixels, and must do so using scalar-wave diffraction from coherent illumination alone. As a result, computer-generated holography (CGH) often has speckling artifacts. We propose HoOPOE (Holography-Optimized Perceptual ObjectivE function), a CGH-specific, perceptually-aware loss function that hides speckles from human vision. HoOPOE is effective for CGH optimizations because it accounts for the human visual system's sensitivity to spatio-temporal frequencies and color, and its parameters are tuned to maximize its performance as a loss function for CGH. We use HoOPOE to generate holograms for mesh-derived focal stacks. We then conduct a user study to benchmark HoOPOE-generated holograms against existing loss functions. HoOPOE yields the most visually pleasing results; users prefer HoOPOE 68% of the time over the next best objective function on average.

cs.GR↗

Specialists Learn Faster: Knowledge Organization in Pre-training Shapes Knowledge Update Efficiency

An ideal (language) model must continually update its knowledge. Existing work largely focuses on making these updates efficient after training. We investigate how the organization of time-varying knowledge during pre-training shapes the efficiency of subsequent updates. Models trained on chronologically ordered knowledge specialize in the most recently encountered knowledge; we call them specialists. In contrast, models trained on the same knowledge in randomly shuffled order perform well across all represented time periods; we call them generalists. Surprisingly, specialists require far fewer update steps than generalists to acquire new knowledge. To explain this advantage, we derive a law from first principles that relates update cost to the model's optimization dynamics and that closely predicts empirical results. Guided by these findings, we design a strategy for organizing pre-training data that preserves stable historical knowledge while enabling efficient updates to current knowledge. We also show that the benefits of temporal knowledge organization extend to real-world pre-trained models and existing knowledge update tools, and generalize to other knowledge orderings. Project page: https://horizon-lab.org/specialist/

cs.LG↗

Multidimensional Observer Model and Perceptual Dimensions of Human Image Quality Assessment

Judging image quality is not only ecologically relevant to everyday human tasks, but also underpins many machine vision tasks such as image generation. This paper proposes a framework to understand the inherent perceptual space underlying image quality judgment in humans. We propose a multi-dimensional observer model that represents images as distributions in a latent perceptual space and that models human judgment as comparing noisy samples. Being constrained by neural representations in the primate ventral stream and fit to large-scale behavioral data, the model enables analysis of perceptual structure while matching the predictive power of existing metrics. Using this model, we find that the perceptual spaces needed to account for image quality judgment in humans are extremely low-dimensional compared to the image space even when considering its sparsity. The exact structure of the space (e.g., dimensionalities, information encoded) varies between low-level and high-level quality judgments, suggesting that, despite a shared retinal encoding in the beginning, humans selectively construct task-dependent perceptual spaces in visual decision making.

cs.CV↗

GazeFlow: From Human Gaze Behavior to Generative Egocentric Gaze Prediction

Egocentric gaze prediction enables many downstream applications but remains challenging, as human gaze is inherently stochastic. This stochasticity is constrained by structured temporal dynamics alternating between fixations and saccades, top-down influences from tasks, and bottom-up visual saliency. Based on this observation, we introduce GazeFlow, a framework that directly models gaze as a joint distribution of temporal gaze positions conditioned upon both top-down and bottom-up information. In particular, GazeFlow uses conditional flow matching (CFM): a learned velocity field iteratively transports a Gaussian noise sample into a plausible gaze trajectory drawn from this joint distribution. The velocity field is conditioned on bottom-up visual features extracted by a video encoder and on top-down task information obtained by globally querying these features. On standard datasets, GazeFlow achieves state-of-the-art performance on per-frame metrics, and the generated trajectories align better with human gaze temporal dynamics.

cs.CV↗

ControlGS: Conditioning Neural Gaussians for Downstream-Processing-Aware XR Rendering

Extended Reality (XR) users do not directly perceive the output of a rendering engine. Instead, rendered images pass through a post-processing pipeline and the physical display-optics path before reaching the eye. Critically, the exact downstream processing can vary significantly at run time, influenced by, for instance, camera pose and display power budget. Traditional 3DGS methods either implicitly assume that this downstream pipeline preserves image quality or cannot adapt to downstream processing changes. To bridge this gap, we present ControlGS, an XR Gaussian rendering pipeline that optimizes end-to-end visual quality. ControlGS models and integrates the entire downstream processing, between the rendering output and the human eye, into the optimization objective. To adapt to downstream processing at run time, ControlGS dynamically generates Gaussian primitives conditioned upon the downstream processing parameters. Experiments show that ControlGS consistently improves end-to-end post-optics XR quality across different neural Gaussian backbones and datasets, with minimal overhead. Code is available at https://horizon-lab.org/controlgs/.

cs.CV↗

LLM-PRISM: Characterizing Silent Data Corruption from Permanent GPU Faults in LLM Training

Large-scale LLM training is increasingly susceptible to hardware defects stemming from manufacturing escapes and silicon aging. These defects manifest as Silent Data Corruption (SDC) that perturb gradients and parameters throughout the training process. We present LLM-PRISM, a methodology to characterize LLM pre-training resilience to hardware faults. LLM-PRISM couples RTL-level GPU fault simulation with a stochastic injection engine embedded in Megatron-LM. Through 7,664 training runs across FP16, BF16, and FP8 regimes, we analyze how fault type, rate, and numeric format govern resilience. We find that while LLMs resist low-frequency faults, impact is highly non-uniform; critical datapaths and specific precision formats can induce catastrophic divergence even at moderate fault rates. This study provides the first hardware-grounded, pre-training characterization of SDC resilience.

cs.AR↗

ControlHair: Synergizing Physics Simulator and Video Diffusion for Controllable Dynamic Hair Rendering

Hair simulation and rendering are challenging due to complex strand dynamics, diverse material properties, and intricate light-hair interactions. Recent video diffusion models can generate high-quality videos, but they lack fine-grained control over hair dynamics. We present ControlHair, a hybrid framework that integrates a physics simulator with conditional video diffusion to enable precise and controllable dynamic hair rendering. ControlHair adopts a three-stage pipeline: it first encodes physics conditions into per-frame geometry using a simulator, then extracts per-frame control signals, and finally feeds control signals into a video diffusion model to generate videos with desired hair dynamics. This cascaded design decouples physics reasoning from video generation, supports diverse physics, and makes training the video diffusion model easy. Trained on a curated 10K video dataset, ControlHair outperforms text- and pose-conditioned baselines, delivering precisely controlled hair dynamics. We also demonstrate three use cases of ControlHair, including dynamic hairstyle try-on, bullet-time effects, and cinemagraphic. Project page: https://linwk20.github.io/controlhair-web.

cs.GR↗

SHUFFLESPARSE: Learned Shuffles for Structured Sparse Networks

Structured weight sparsity accelerates training and inference on modern GPUs, but it trails unstructured dynamic sparse training (DST) in accuracy especially at extreme sparsity. We pinpoint the reason for this difference in performance to a lack of expressivity: a dense layer can implement any pattern of non-zero weights, whereas structured patterns are restricted to only a small set of weight configurations. We introduce SHUFFLESPARSE, a single permutation primitive that applies uniformly across DST-from-scratch and one-shot pruning, and across N:M, block etc. We close most of this gap by learning a single permutation matrix jointly with the structured weight matrix. When used on three different types of structures (block, N:M, and diagonal), SHUFFLESPARSE is able to reduce the structured-vs-unstructured accuracy gap on ViT-B16 (ImageNet-1K) and GPT-2 (WikiText-103) at 90-95% sparsity, while adding minimal inference overhead (< 8.7% inference overhead) and preserving any training acceleration the host structure provides. The same permutation formulation transfers to one-shot 2:4 pruning of pretrained LLMs, where it improves zero-shot accuracy by 4.6 points on LLaMA-2 7B. Together, these results establish learned permutations as a general tool for recovering unstructured-level accuracy from structured sparsity patterns.

cs.LG↗

LowPowAR: Power-Constrained Tone Mapping for Augmented Reality

Everyday-wearable Augmented Reality (AR) glasses must meet strict power limits, making displays a key target for optimization. We cast display power optimization as a power-constrained tone-mapping problem and propose a human-vision-grounded, learning-based framework that maximizes perceptual quality under a given power budget. We introduce an optimization-friendly tone-mapping operator (TMO) parameterization along with a progressive optimization strategy to effectively navigate the quality-vs-power landscape. We distill the iterative optimization into a lightweight feed-forward neural network for real-time deployment. Subjective experiments show that our method yields better perceptual quality than prior work at the same power budget. Project page: https://horizon-lab.org/lowpowar/.

cs.GR↗

Flux Variations of Fast Radio Bursts and Their Persistent Radio Sources: Evidence for a Shared Progenitor

Fast radio bursts (FRBs) are millisecond-duration extragalactic radio transients, some of which are associated with compact persistent radio sources (PRSs), hinting at a physical connection. While several models have been proposed to explain PRSs and their connection to FRBs, direct observational tests remain limited. Here, we report for the first time a correlated trend between the long-term variation of the PRS flux density and the burst energetics of FRB 20190520B and FRB 20240114A, suggesting a physical coupling between the PRS and FRB activity. We further examine additional repeaters with compact PRSs and find no clear correlation between PRS luminosity and burst activity, likely due to the limited observations. These results are consistent with scenarios in which both the PRS and FRB activity may be powered by a common energy reservoir, such as the magnetic or rotational energy of a magnetar.

astro-ph.HE↗

A Persistently Active Fast Radio Burst source Embedded in an Expanding Supernova Remnant

Fast radio bursts (FRBs) remain one of the most puzzling astrophysical phenomena. While most FRBs are detected only once or sporadically, we present the identification of FRB 20190520B as the first persistently active source over a continuous span of ~ four years. This rare long-term activity enabled a detailed investigation of its dispersion measure (DM) evolution. We also report that FRB 20190520B exhibits a substantial decrease in DM at a global rate of minus 12.4 plus or minus 0.3 pc cm^-3 yr^-1, exceeding previous FRB DM variation measurements by a factor of three and surpassing those observed in pulsars by orders of magnitude. The magnitude and consistency of the DM evolution, along with a high host DM contribution, strongly indicate that the source resides in a dense, expanding ionized medium, likely a young supernova remnant (SNR).

astro-ph.HE↗

Investigating FRB 20240114A with FAST: Morphological Classification and Drifting Rate Measurements in a Burst-Cluster Framework

This study investigates the morphological classification and drifting rate measurement of the repeating fast radio burst (FRB) source FRB 20240114A using the Five-hundred-meter Aperture Spherical Telescope (FAST). Detected on January 14, 2024, FRB 20240114A exhibited an exceptionally high burst rate, revealing unique properties. Through observational campaigns over several months, we selected a dataset comprising 3,203 bursts (2,109 burst-clusters) during a continuous monitoring session (15,780 seconds) on March 12, 2024. Improving upon previous work, we clarify the definitions of sub-bursts, bursts and burst-clusters. Using an average dispersion measures (DM) of 529.2 pc cm$^{-3}$, we classified the burst-clusters into Downward Drifting, Upward Drifting, No Drifting, No Evidence for Drifting, Not-Clear, and Complex burst-clusters. Among the 978 burst-clusters that exhibit drifting behavior, 233 (23.82%) show upward drifting. Additionally, if 142 upward drifting single-component burst-clusters are excluded, upward drifting double- and multi-component burst-clusters still account for 10.89% of the 836 burst-clusters exhibiting drifting behavior, equating to 91 burst-clusters. Furthermore, if only upward drifting burst-clusters with consecutive time intervals (or upward drifting bursts) are considered, only 9 bursts remain. Drifting rate comparisons with other physical quantities reveal that the drifting rate increases with peak frequency for single-component burst-clusters with drifting behavior. Moreover, in single-component burst-clusters, those with upward drifting exhibit smaller effective widths, bandwidths, and fluxes than their downward drifting counterparts. A Kolmogorov-Smirnov test further indicates that upward drifting burst-clusters possess longer consecutive time intervals than downward drifting ones, suggesting distinct underlying physical mechanisms.

astro-ph.HE↗

A comprehensive search for Long and Short Periodic Features from an Extremely Active Cycle of FRB 20240114A

Possible periodic features in fast radio bursts (FRBs) may provide insights into their astrophysical origins. Using extensive observations from the Five-hundred-meter Aperture Spherical radio Telescope (FAST), we conduct a multi-timescale periodicity search for the exceptionally active repeater FRB~20240114A. Our analysis is based on different datasets for different timescales: for short-timescale periodicity in Time of Arrivals (TOAs), we use 57 observations from January to August 2024; for long-timescale periodicity, we employ an extended TOA dataset comprising 111 observations spanning from January 2024 to October 2025; and for burst time series analysis, we utilize individual burst data from the 57 FAST observations. We identify three candidate short-timescale periodic signals (0.673~s, 0.635~s, and 0.536~s) with significances of $3.2σ$--$6σ$, each detected in two independent observations. On longer timescales, we detect a significant $143.40\pm7.19$-day periodicity with $5.2σ$ significance, establishing FRB~20240114A as a periodic repeater. In burst time series, we find quasi-periodic oscillations in the few hundred Hz range ($3.4σ$ and $3.7σ$) and periodic burst trains with periods of several to tens of milliseconds ($3σ$--$3.9σ$), though these periodic features appear transient and short-lived. The detection of periodic signals at these different time scales indicates that FRB 20240114A exhibits intriguing periodic self-similar characteristics. Despite the comprehensive dataset, no definitive periodicity linked to the source's rotation is confirmed, placing stringent constraints on the intrinsic source properties and the modulation mechanisms. All data are available via the Science Data Bank.

astro-ph.HE↗

Revealing the Temporally Stable Bimodal Energy Distribution of FRB 20121102A with a Tripled Burst Set from AI Detections

Active repeating Fast Radio Bursts (FRBs), with their large number of bursts, burst energy distribution, and their potential energy evolution, offer critical insights into the FRBs emission mechanisms. Traditional pipelines search for bursts through conducting dedispersion trials and looking for signals above certain fluence thresholds, both of which could result in missing weak and narrow-band bursts. In order to improve the completeness of the burst set, we develop an End-to-end DedispersE-agnostic Nonparametric AI model (EDEN), which directly detect bursts from dynamic spectrum and is the first detection pipeline that operates without attempting dedispersion. We apply EDEN to archival FAST L-band observations during the extreme active phase of the repeating source FRB 20121102A, resulting in the largest burst set for any FRB to date, which contains 5,927 individual bursts, tripling the original burst set. The much enhanced completeness enables a refined analysis of the temporal behavior of energy distribution, revealing that the bimodal energy distribution remains stable over time. It is rather an intrinsic feature of the emission mechanisms than a consequence of co-evolving with burst rate.

astro-ph.HE↗

Modeling and Exploiting the Time Course of Chromatic Adaptation for Display Power Optimizations in Virtual Reality

We introduce a gaze-tracking--free method to reduce OLED display power consumption in VR with minimal perceptual impact. This technique exploits the time course of chromatic adaptation, the human visual system's ability to maintain stable color perception under changing illumination. To that end, we propose a novel psychophysical paradigm that models how human adaptation state changes with the scene illuminant. We exploit this model to compute an optimal illuminant shift trajectory, controlling the rate and extent of illumination change, to reduce display power under a given perceptual loss budget. Our technique significantly improves the perceptual quality over prior work that applies illumination shifts instantaneously. Our technique can also be combined with prior work on luminance dimming to reduce display power by 31% with no statistical loss of perceptual quality.

cs.GR↗