arXiv subjects

Liang

Publications and source records attributed to Liang.

9 recordsLinked to original sources

arXiv2026

SPADE: An Input-Adaptive Sparse Attention Engine for Fast Video Diffusion Models Inference

Video diffusion transformers (vDiTs) generate high quality but pay quadratic self-attention cost, making inference prohibitive at video-token scales. The challenge is input-adaptive sparsity: selecting critical Q/K/V tokens with negligible overhead and executing them for end-to-end gains. We present SPADE, a training-free sparse-attention engine of three parts: (i) vDiT-SSR, a specification defining 3D blocking candidates and formalizing dynamic masks via Summarizer/Estimator expressions; (ii) runtime scheme generation using SICS and a head-wise policy; and (iii) an executor with low-overhead index search, flash block-sparse attention, and kernel grouping. Across Hunyuan-Video and Wan 2.1/2.2 for text-to-video and image-to-video generation, SPADE raises sparsity and speed while preserving quality, accelerating attention by 2.26x-3.40x and end-to-end inference by 1.49x-1.80x. Our code is open-sourced at https://github.com/6somehow/DAC-SPADE.

cs.CV↗

arXiv2026

Moment Matching Q-Learning

Score-based and flow-based generative models exhibit remarkable expressive capacity in capturing complex distributions, and have been extensively deployed in tasks ranging from image generation to reinforcement learning. Nevertheless, these models suffer from prolonged inference latency, which imposes a significant computational bottleneck in RL with iterative sampling. To overcome this limitation, we propose a new framework named Moment Matching Q-Learning (MoMa QL), which utilizes a technique from statistical hypothesis testing known as maximum mean discrepancy (MMD) that intend to match all orders of statistics between the original and target distribution. By enforcing strong regularization on all moment statistics, this algorithm guarantees distribution-level convergence for conditional score function and remains stable under various hyperparameters. Empirically, we show that our method MoMa QL is more computationally efficient with a comparable if not competitive performance in various D4RL tasks. Remarkably, by accelerating the action sampling process for flow-based policies, MoMa QL demonstrates superior performance in offline-to-online RL tasks because of faster and stronger adaptability for online interactive finetuning.

cs.LG↗

arXiv2026

AI-Mediated Hiring and the Job Search of Blind and Low-Vision Individuals

Blind and low-vision (BLV) individuals face high unemployment rates. The job search is becoming harder as more employers use AI-driven systems to screen resumes before a human ever sees them. Such AI systems could inadvertently further disadvantage BLV job seekers, introducing additional barriers to an already difficult process. We lack understanding of BLV job seekers' experiences in today's AI-driven hiring ecosystem. Without such understanding, we risk designing technologies that create new systemic barriers for BLV job seekers rather than providing support. To this end, we conducted interviews with 17 BLV job seekers and analyzed their experiences with AI-powered hiring systems. We found that AI hiring systems misrepresented their professional identities and created dehumanizing interactions. To level the playing field, BLV job seekers used strategic counter-navigation: they deployed their own tools to bypass algorithmic screening and built peer networks to share AI literacy. They also practiced 'strategic refusal', choosing to avoid certain AI systems to regain their agency. Unlike prior work that frames job search as an individualistic activity, or one focused on being compliant with employer needs, we use the interdependence framework to argue that for BLV people, job search is an interdependent process. We offer design recommendations for AI-mediated tools that center disability perspectives and support interdependencies in job search.

cs.HC↗

arXiv2025

NOVA: Coordinated Test Selection and Bayes-Optimized Constrained Randomization for Accelerated Coverage Closure

Functional verification relies on large simulation-based regressions. Traditional test selection relies on static test features and overlooks actual coverage behavior, wasting substantial simulation time, while constrained random stimuli generation depends on manually crafted distributions that are difficult to design and often ineffective. We present NOVA, a framework that coordinates coverage-aware test selection with Bayes-optimized constrained randomization. NOVA extracts fine-grained coverage features to filter redundant tests and modifies the constraint solver to expose parameterized decision strategies whose settings are tuned via Bayesian optimization to maximize coverage growth. Across multiple RTL designs, NOVA achieves up to a 2.82$\times$ coverage convergence speedup without requiring human-crafted heuristics.

stat.ME↗

arXiv2025

FedHQ: Hybrid Runtime Quantization for Federated Learning

Federated Learning (FL) is a decentralized model training approach that preserves data privacy but struggles with low efficiency. Quantization, a powerful training optimization technique, has been widely explored for integration into FL. However, many studies fail to consider the distinct performance attribution between particular quantization strategies, such as post-training quantization (PTQ) or quantization-aware training (QAT). As a result, existing FL quantization methods rely solely on either PTQ or QAT, optimizing for speed or accuracy while compromising the other. To efficiently accelerate FL and maintain distributed convergence accuracy across various FL settings, this paper proposes a hybrid quantitation approach combining PTQ and QAT for FL systems. We conduct case studies to validate the effectiveness of using hybrid quantization in FL. To solve the difficulty of modeling speed and accuracy caused by device and data heterogeneity, we propose a hardware-related analysis and data-distribution-related analysis to help identify the trade-off boundaries for strategy selection. Based on these, we proposed a novel framework named FedHQ to automatically adopt optimal hybrid strategy allocation for FL systems. Specifically, FedHQ develops a coarse-grained global initialization and fine-grained ML-based adjustment to ensure efficiency and robustness. Experiments show that FedHQ achieves up to 2.47x times training acceleration and up to 11.15% accuracy improvement and negligible extra overhead.

cs.LG↗

arXiv2024

OriGen:Enhancing RTL Code Generation with Code-to-Code Augmentation and Self-Reflection

Recent studies have demonstrated the significant potential of Large Language Models (LLMs) in generating Register Transfer Level (RTL) code, with notable advancements showcased by commercial models such as GPT-4 and Claude3-Opus. However, these proprietary LLMs often raise concerns regarding privacy and security. While open-source LLMs offer solutions to these concerns, they typically underperform commercial models in RTL code generation tasks, primarily due to the scarcity of high-quality open-source RTL datasets. To address this challenge, we introduce OriGen , a fully open-source framework that incorporates self-reflection capabilities and a novel dataset augmentation methodology for generating high-quality, large-scale RTL code. Our approach employs a code-tocode augmentation technique to enhance the quality of open-source RTL code datasets. Furthermore, OriGen can rectify syntactic errors through a self-reflection process that leverages compiler feedback. Experimental results demonstrate that OriGen significantly outperforms other open-source alternatives in RTL code generation. It surpasses the previous best-performing open-source LLM by 12.8% and even exceeds GPT-4 Turbo in the pass@1 metric on the VerilogEval-Human benchmark. Moreover, OriGen exhibits superior capabilities in self-reflection and error correction, outperforming GPT-4 by 19.9% on a benchmark designed to evaluate self-reflection capabilities.

cs.AR↗

arXiv2024

Implementing Bayesian inference on a stochastic CO2-based grey-box model for assessing indoor air quality in Canadian primary schools

The COVID-19 pandemic brought global attention to indoor air quality (IAQ), which is intrinsically linked to clean air change rates. Estimating the air change rate in indoor environments, however, remains challenging. It is primarily due to the uncertainties associated with the air change rate estimation, such as pollutant generation rates, dynamics including weather and occupancies, and the limitations of deterministic approaches to accommodate these factors. In this study, Bayesian inference was implemented on a stochastic CO2-based grey-box model to infer modeled parameters and quantify uncertainties. The accuracy and robustness of the ventilation rate and CO2 emission rate estimated by the model were confirmed with CO2 tracer gas experiments conducted in an airtight chamber. Both prior and posterior predictive checks (PPC) were performed to demonstrate the advantage of this approach. In addition, uncertainties in real-life contexts were quantified with an incremental variance {\sigma} for the Wiener process. This approach was later applied to evaluate the ventilation conditions within two primary school classrooms in Montreal. The Equivalent Clean Airflow Rate (ECAi) was calculated following ASHRAE 241, and an insufficient clean air supply within both classrooms was identified. A supplement of 800 cfm clear air delivery rate (CADR) from air-cleaning devices is recommended for a sufficient ECAi. Finally, steady-state CO2 thresholds (Climit, Ctarget, and Cideal) were carried out to indicate when ECAi requirements could be achieved under various mitigation strategies, such as portable air cleaners and in-room ultraviolet light, with CADR values ranging from 200 to 1000 cfm.

stat.AP↗

arXiv2019

On the Optical Accuracy of the Salvator Mundi

A debate in the scientific literature has arisen regarding whether the orb depicted in Salvator Mundi, which has been attributed by some experts to Leonardo da Vinci, was rendered in a optically faithful manner or not. Some hypothesize that it was solid crystal while others hypothesize that it was hollow, with competing explanations for its apparent lack of background distortion and its three white spots. In this paper, we study the optical accuracy of the Salvator Mundi using physically based rendering, a sophisticated computer graphics tool that produces optically accurate images by simulating light transport in virtual scenes. We created a virtual model of the composition centered on the translucent orb in the subject's hand. By synthesizing images under configurations that vary illuminations and orb material properties, we tested whether it is optically possible to produce an image that renders the orb similarly to how it appears in the painting. Our experiments show that an optically accurate rendering qualitatively matching that of the painting is indeed possible using materials, light sources, and scientific knowledge available to Leonardo da Vinci circa 1500. We additionally tested alternative theories regarding the composition of the orb, such as that it was a solid calcite ball, which provide empirical evidence that such alternatives are unlikely to produce images similar to the painting, and that the orb is instead hollow.

cs.GR↗

arXiv2012

Dynamics of unconfined spherical flames

Using the soap bubble technique, we visualize the dynamics of unconfined hydrogen-air flames using high speed schlieren video. We show that for sufficiently weak mixtures, i.e., low flame speeds, buoyancy effects become important. Flame balls of a critical dimension begin to rise. The experiments are found in very good agreement with the scaling laws proposed by Zingale and Dursi. We report the results in a fluid dynamics video.

physics.flu-dyn↗