arXiv subjects

Qicheng Zhao

Publications and source records attributed to Qicheng Zhao.

4 recordsLinked to original sources

arXiv2026

Seeing the End at Step Zero: Accelerating Diffusion MLLMs via MLP Sparsity-Aware Truncation

Diffusion Multimodal Large Language Models (DMLLMs) are highly effective for multimodal reasoning, yet their inference efficiency is significantly hindered by fixed-length generation constraints. Since the actual output length is unknown, output sequences are padded to a predefined maximum length, resulting in substantial redundant computation over unnecessary [EOS] tokens. In this work, we discover that DMLLMs implicitly reveal their valid semantic boundary at the very first denoising step through a distinct shift in MLP activation sparsity. Leveraging this observation, we propose Seer, a training-free framework that detects this boundary using a Signal-to-Noise Ratio (SNR)-based criterion and performs one-shot truncation of the redundant suffix for all subsequent computations. To preserve these theoretical gains during batched serving, Seer incorporates a hybrid execution strategy that maximizes throughput while seamlessly accommodating dynamic sequence lengths. Experimental results demonstrate that Seer effectively eliminates padding waste, accelerating throughput by up to $\sim$31$\times$. Across 9 benchmarks, Seer robustly maintains overall performance and even improves accuracy on complex visual tasks by mitigating noise leakage (e.g., DocVQA score increases from 63.52 to 63.66), offering a highly efficient, plug-and-play solution for DMLLM acceleration.

cs.AI↗

arXiv2026

ResilPhase: Plug-and-Play Phase Mapping and Noise-Resilient Macro-Trajectory Extrapolation for Diffusion Acceleration

The adoption of powerful diffusion models is hindered by their significant inference latency. Recent ``cache-then-forecast'' schemes alleviate this issue by accelerating DiTs using derivative-based polynomials, but they suffer from severe quality degradation at high acceleration ratios. Our analysis reveals its root cause: the discrete extrapolation performed on representations that are misaligned with the continuous diffusion trajectory and are numerically unstable. Thus, accelerated DiTs suffer from accumulated spatial errors, noisy derivative amplification, and high-order instability. We therefore reformulate accelerated inference as stable macro-trajectory extrapolation in ordinary differential equation (ODE) space. Instead of predicting intermediate features, we align forecasting with the model's Global Drift (GD), i.e., the end-to-end state evolution, thereby eliminating feature inconsistency and memory overhead. However, even this smooth macro-trajectory remains vulnerable to the derivative fallacy: its higher-order temporal derivatives are intrinsically noisy. Thus, we introduce a derivative-free barycentric Lagrange extrapolator to effectively bypass derivative instability and approximation error. We further propose a bounded Phase Mapping that regularizes the extrapolation domain, suppressing oscillatory error growth. These elements collectively constitute ResilPhase, a noise-resilient acceleration framework. Experiments on FLUX.1-dev and HunyuanVideo demonstrate state-of-the-art fidelity under aggressive acceleration ratios.

cs.AI↗

arXiv2026

Varying-coefficient mixture of experts model for dynamic heterogeneous populations: application to mouse cortical development

As cells differentiate, gene-gene associations may change. Because the composition of cell subtypes also shifts with development, it is challenging to establish whether those changes reflect real changes in gene regulation within one or more subpopulations or a spurious one due to the shifting composition. Only formal inference can separate them. We propose a Varying-Coefficient Mixture-of-Experts (VCMoE) model in which the coefficients of both the gating functions and the expert components vary smoothly with the index variable, so that a subgroup-specific association can be inferred without being confounded by the concurrent shift in subgroup composition. We develop a label-consistent EM algorithm and establish the identifiability and asymptotic theory that make the resulting inference valid. These guarantees support simultaneous confidence bands, constructed using asymptotic and bootstrap methods, together with a generalized likelihood ratio test for constant coefficient functions. Simulations demonstrate accurate estimation and satisfactory coverage. Applied to single-nucleus RNA-sequencing data from embryonic mouse cortex, VCMoE finds that the repression of Bcl11b by Satb2 within the upper layer is not yet in place when those neurons first appear and becomes established over the course of differentiation, a dynamic that an ordinary mixture-of-experts model neglected. A package for implementing the method is available.

stat.ME↗

arXiv2024

High-dimensional Bayesian Model for Disease-Specific Gene Detection in Spatial Transcriptomics

Identifying disease-indicative genes is critical for deciphering disease mechanisms and has attracted significant interest in biomedical research. Spatial transcriptomics offers unprecedented insights for the detection of disease-specific genes by enabling within-tissue contrasts. However, this new technology poses challenges for conventional statistical models developed for RNA-sequencing, as these models often neglect the spatial organization of tissue spots. In this article, we propose a Bayesian shrinkage model to characterize the relationship between high-dimensional gene expressions and the disease status of each tissue spot, incorporating spatial correlation among these spots through autoregressive terms. Our model adopts a hierarchical structure to facilitate the analysis of multiple correlated samples and is further extended to accommodate the missing data within tissues. To ensure the model's applicability to datasets of varying sizes, we carry out two computational frameworks for Bayesian parameter estimation, tailored to both small and large sample scenarios. Simulation studies are conducted to evaluate the performance of the proposed model. The proposed model is applied to analyze the data arising from a HER2-positive breast cancer study.

stat.ME↗