arXiv subjects

Xinxi Zhang

Publications and source records attributed to Xinxi Zhang.

5 recordsLinked to original sources

arXiv2026

Few-Step Diffusion Language Models via Trajectory Self-Distillation

Diffusion large language models (DLLMs) have emerged as powerful generative models with the promise of fast text generation through parallel decoding. However, realizing this potential in practice remains challenging: reducing the number of decoding steps, typically causes a substantial degradation in output quality due to token factorization error. To alleviate this, we propose a self-distillation framework that trains a few-step student to match the generative trajectory of a full-step teacher. We theoretically and empirically show that trajectory-level supervision mitigates this factorization error, thereby enabling effective few-step decoding. We further incorporate Direct Discriminative Optimization (DDO), a reverse-KL objective that encourages mode-seeking toward the teacher's modes, yielding stronger performance on challenging reasoning tasks. Across reasoning and code-generation benchmarks, our method substantially narrows the gap between few-step and full-step decoding. The source code is available at https://github.com/Tyrion58/T3D.

cs.CL↗

arXiv2025

Overcoming the Curvature Bottleneck in MeanFlow

MeanFlow offers a promising framework for one-step generative modeling by directly learning a mean-velocity field, bypassing expensive numerical integration. However, we find that the highly curved generative trajectories of existing models induce a noisy loss landscape, severely bottlenecking convergence and model quality. We leverage a fundamental geometric principle to overcome this: mean-velocity estimation is drastically simpler along straight paths. Building on this insight, we propose Rectified MeanFlow, a self-distillation approach that learns the mean-velocity field over a straightened velocity field, induced by rectified couplings from a pretrained model. To further promote linearity, we introduce a distance-based truncation heuristic that prunes residual high-curvature pairs. By smoothing the optimization landscape, our method achieves strong one-step generation performance. We improve the FID of baseline MeanFlow models from 30.9 to 8.6 under same training budget, and outperform the recent 2-rectified flow++ by 33.4% in FID while running 26x faster. Our work suggests that the difficulty of one-step flow generation stems partially from the rugged optimization landscapes induced by curved trajectories. Code is available at https://github.com/Xinxi-Zhang/Re-MeanFlow.

cs.CV↗

arXiv2024

DeepMIDE: A Multi-Output Spatio-Temporal Method for Ultra-Scale Offshore Wind Energy Forecasting

To unlock access to stronger winds, the offshore wind industry is advancing towards significantly larger and taller wind turbines. This massive upscaling motivates a departure from wind forecasting methods that traditionally focused on a single representative height. To fill this gap, we propose DeepMIDE--a statistical deep learning method which jointly models the offshore wind speeds across space, time, and height. DeepMIDE is formulated as a multi-output integro-difference equation model with a multivariate nonstationary kernel characterized by a set of advection vectors that encode the physics of wind field formation and propagation. Embedded within DeepMIDE, an advanced deep learning architecture learns these advection vectors from high-dimensional streams of exogenous weather information, which, along with other parameters, are plugged back into the statistical model for probabilistic multi-height space-time forecasting. Tested on real-world data from offshore wind energy areas in the Northeastern United States, the wind speed and power forecasts from DeepMIDE are shown to outperform those from prevalent time series, spatio-temporal, and deep learning methods.

cs.LG↗

arXiv2024

Spectrum-Aware Parameter Efficient Fine-Tuning for Diffusion Models

Adapting large-scale pre-trained generative models in a parameter-efficient manner is gaining traction. Traditional methods like low rank adaptation achieve parameter efficiency by imposing constraints but may not be optimal for tasks requiring high representation capacity. We propose a novel spectrum-aware adaptation framework for generative models. Our method adjusts both singular values and their basis vectors of pretrained weights. Using the Kronecker product and efficient Stiefel optimizers, we achieve parameter-efficient adaptation of orthogonal matrices. We introduce Spectral Orthogonal Decomposition Adaptation (SODA), which balances computational efficiency and representation capacity. Extensive evaluations on text-to-image diffusion models demonstrate SODA's effectiveness, offering a spectrum-aware alternative to existing fine-tuning methods.

cs.CV↗

arXiv2023

PAC Privacy Preserving Diffusion Models

Data privacy protection is garnering increased attention among researchers. Diffusion models (DMs), particularly with strict differential privacy, can potentially produce images with both high privacy and visual quality. However, challenges arise such as in ensuring robust protection in privatizing specific data attributes, areas where current models often fall short. To address these challenges, we introduce the PAC Privacy Preserving Diffusion Model, a model leverages diffusion principles and ensure Probably Approximately Correct (PAC) privacy. We enhance privacy protection by integrating a private classifier guidance into the Langevin Sampling Process. Additionally, recognizing the gap in measuring the privacy of models, we have developed a novel metric to gauge privacy levels. Our model, assessed with this new metric and supported by Gaussian matrix computations for the PAC bound, has shown superior performance in privacy protection over existing leading private generative models according to benchmark tests.

cs.LG↗