Selected work

Representative Papers

GenzIQA: Generalized Image Quality Assessment using Prompt-Guided Latent Diffusion Models

Jun 07, 2024arXiv.org

Existing no-reference image quality assessment (IQA) methods exhibit poor cross-dataset generalization, particularly under distribution shifts such as user-generated content, synthetic imagery, and low-light conditions. To address this, we propose the first generic IQA framework leveraging the cross-attention mechanism of text-guided latent diffusion models (LDMs). Our method introduces learnable, quality-aware textual prompts and models prompt–image alignment to derive robust quality representations. Crucially, it exploits intermediate cross-attention features from the LDM denoising process—enabling zero-shot transfer to multiple benchmark datasets without fine-tuning. Extensive experiments demonstrate that our approach significantly outperforms state-of-the-art methods on diverse databases including LIVE-Youtube, KoNViD, and UHD-1. Moreover, it achieves superior out-of-distribution generalization, validating its effectiveness under substantial domain shifts. This work establishes a novel paradigm for leveraging generative model priors in blind IQA, bridging semantic understanding and perceptual quality estimation.

5 citations1 influentialRead paper

QUOKA: Query-Oriented KV Selection For Efficient LLM Prefill

Feb 09, 2026

This work addresses the high computational cost and inefficiency of dense attention during the prefilling phase in large language model inference. The authors propose a training-free, hardware-agnostic sparse attention algorithm that, for the first time, leverages the observation that queries with low cosine similarity dominate attention outcomes. By employing a query-guided strategy to select the most representative queries and their best-aligned key-value pairs, the method efficiently approximates full attention. Evaluated across multiple benchmarks, the approach achieves accuracy nearly on par with the dense baseline while reducing key-value pair usage by 88%. It delivers a 5× speedup on NVIDIA GPUs and nearly 7× on Intel Xeon CPUs, along with a 3× reduction in first-token latency.

1 citationsRead paper

Direct Preference Optimization with Rating Information: Practical Algorithms and Provable Gains

Jan 31, 2026

This work addresses a key limitation of existing Direct Preference Optimization (DPO) methods, which rely solely on pairwise preference signals and neglect the quantitative differences in response quality, leading to ambiguous training signals and suboptimal optimization efficiency. To overcome this, we propose a novel preference optimization algorithm that, for the first time, incorporates explicit score gaps into the DPO framework. Our approach preserves the advantage of not requiring an explicit reward model while leveraging fine-grained relative quality information to enhance alignment. By designing a loss function that accounts for score differences, the method enjoys faster theoretical statistical convergence and demonstrates robustness to scoring noise. Extensive experiments show consistent and significant improvements over current DPO variants across multiple large language models and evaluation benchmarks, with stable performance gains even when provided with inaccurate scores.

1 citationsRead paper

ReHyAt: Recurrent Hybrid Attention for Video Diffusion Transformers

Jan 07, 2026arXiv.org

This work addresses the challenge of efficiently generating long-sequence videos with transformer-based video diffusion models, which are hindered by the quadratic computational complexity of standard attention mechanisms. To overcome this limitation, the authors propose a recurrent hybrid attention mechanism that integrates the high-fidelity modeling capacity of softmax attention with the computational efficiency of linear attention. This design enables chunk-wise recurrent modeling with constant memory consumption and facilitates efficient knowledge distillation from existing pretrained models. The proposed method reduces attention complexity from O(n²) to O(n) and achieves state-of-the-art video generation quality on VBench, VBench-2.0, and human evaluations, while significantly lowering training costs to approximately 160 GPU hours.

1 citationsRead paper

Optimizing System Latency for Blockchain-Encrypted Edge Computing in Internet of Vehicles

Jun 17, 2025Computers, Materials & Continua

To address the dual challenges of adversarial external attacks and high latency induced by blockchain integration in Internet of Vehicles (IoV) edge task offloading, this paper proposes a secure, low-latency edge computing framework incorporating the Raft consensus mechanism. We pioneer the integration of a lightweight Raft protocol into the IoV edge architecture to establish a tamper-resistant and verifiable task offloading security mechanism. An end-to-end latency analytical model is developed to characterize the coupled delays arising from communication, computation, and consensus. Subsequently, a convex-optimization-based joint resource-consensus scheduling algorithm is designed to achieve Pareto-optimal trade-offs between security and latency. Simulation results demonstrate that, compared with baseline schemes, the proposed framework reduces average system latency by 32.7%, decreases latency standard deviation by 41.5%, and achieves a data extraction rate of 99.2%, thereby satisfying IoV requirements for millisecond-level responsiveness and high reliability.

1 citationsRead paper

Recent publications

Latest Papers

Semantic Watermarking with Order-Robust Detection over Sub-sentence Units

Aug 27, 2026

本文针对语义水印易受重写、重组和重新分段攻击的问题，提出了一种基于子句单元的顺序鲁棒检测方法k-SwordStamp。

0 citationsRead paper

MoNe: Modular Neural Memory for Efficient Long Context Inference

Aug 18, 2026

本文提出MoNe，一种轻量级模块化神经记忆体，通过两阶段设计实现长上下文推理，减少计算和GPU内存消耗，提高长文本处理性能。

0 citationsRead paper

MagnifiQ: Patch-aware Text Guided Progressive Upscaling for High-Resolution Image Restoration

Aug 14, 2026

This study addresses the challenges of poor global consistency, repetitive textures, and high computational costs in high-resolution image inpainting by proposing an efficient progressive upsampling framework based on SDXL. The method replaces self-attention layers with linear-complexity convolutions and integrates tiled text semantic guidance with a multi-stage iterative strategy, effectively overcoming computational bottlenecks for 4K restoration. Experimental results demonstrate that the proposed model significantly outperforms existing methods in both perceptual quality and human preference, efficiently generating clear and coherent high-resolution images. Consequently, this work achieves a superior balance between restoration fidelity and computational efficiency, offering a scalable solution for ultra-high-resolution image editing tasks.

0 citationsRead paper

SNM-VFI: Symmetric Nonlinear Motion-Guided Generative Video Frame Interpolation

Aug 13, 2026

This work addresses the challenges of poor motion controllability, low perceptual quality, and temporal inconsistency in video frame interpolation by proposing a training-free interpolation framework. It leverages a pretrained optical flow model to construct symmetric nonlinear motion-guided frames, which serve as latent-space priors to iteratively steer a pretrained video diffusion model for high-fidelity and motion-coherent intermediate frame synthesis. The method innovatively integrates symmetric nonlinear motion modeling with a pretrained video diffusion model and introduces a confidence map fusion mechanism that balances structural reliability and textural realism in ambiguous regions such as occlusions and object boundaries. Extensive experiments on standard benchmarks—including DAVIS, Sintel, and KITTI—demonstrate superior performance in perceptual quality, reconstruction accuracy, and temporal consistency.

0 citationsRead paper

Gated-BEPO: Confidence-Gated Bellman Credit Assignment for Large Language Model Agents

Aug 07, 2026

This work addresses the challenge of credit assignment in long-horizon sparse-reward environments by proposing a step-level credit assignment method based on an experience replay graph. The approach constructs an experience graph and estimates node values via mean-backed Bellman fixed-point updates, then computes step-level Bellman advantages using generalized advantage estimation. A confidence-gated mechanism adaptively fuses episode-level and step-level credit signals, enabling selective rather than uniform credit assignment across all timesteps. Evaluated on tasks including WebShop, ALFWorld, and visual Sokoban, the method significantly enhances the performance of both language model and vision-language model agents. Ablation studies confirm the effectiveness of each proposed component.

0 citationsRead paper