arXiv subjects

Zhiqiang Tan

Publications and source records attributed to Zhiqiang Tan.

At least 19 recordsLinked to original sources

arXiv2026

LeanGRPO: Eliminating Redundant Recomputation in Diffusion RL

Diffusion reinforcement learning (RL) has recently achieved significant success in post-training image and video generative models. However, most diffusion RL methods, including DanceGRPO and FlowGRPO, recompute selected timesteps with gradient tracking after rollout. Under on-policy training with the same backend for rollout and update, this recomputation is mathematically redundant. Intuitively, the rollout and policy update steps can reuse the same feed-forward backbone to avoid redundant computation, but doing so can incur a large memory overhead during rollout. To address the issue, we present LeanGRPO by restructuring the data-parallel layout and introducing two recompute-free training schedules for trajectory-logprob diffusion RL: (1) LeanGRPO-Retain enables gradient tracking during rollout and directly reuses the resulting computation graphs and saved activations for backward during update, requiring no recomputation; and (2) LeanGRPO-Reweight also enables gradients during rollout, but immediately backpropagates each selected step using a provisional advantage and delays gradient synchronization, then corrects the provisional gradients with the true advantage after the trajectory is completed. These schedules target different model scales and input sizes. Across FlowGRPO/DanceGRPO with FLUX.1-dev and Wan, LeanGRPO achieves up to 1.83x end-to-end speedup while preserving the original optimization objective.

cs.LG↗

arXiv2026

Unregularized Convergence of Single-Loop, Entropy-Regularized Natural Actor-Critic

While entropy regularization is widely used to stabilize and accelerate Natural Policy Gradient methods, its ability to yield faster convergence rates for the unregularized objective remains underexplored. Existing analyses often rely on double-loop architectures and invoke a linear entropy penalty. To bridge the gap between theory and practice, we analyze a single-loop, entropy-regularized Natural Actor-Critic algorithm under compatible linear function approximation. By training an uncentered critic, our critic tracking can remain stable even as the training policy approaches determinism and the Fisher information matrix degenerates. We focus on two primary regimes for the optimization landscape: a Stochastic Regime, where we fuse coupled actor-critic updates into a joint Lyapunov recurrence, and a Deterministic Regime, where we pivot to a Policy Mirror Descent framework to circumvent the collapse of Euclidean geometry. By exploiting a positive Minimal Action Gap in the unregularized Markov decision process, we introduce an Exponential Translation mechanism that maps the regularized gap to the unregularized one up to an exponentially decaying tail. By tuning the fixed temperature, our algorithm achieves accelerated unregularized convergence rates, up to approximation-error terms: $\tilde{\mathcal{O}}(T_{total}^{-1})$ in the Stochastic Regime, and $\tilde{\mathcal{O}}(T_{total}^{-2/3})$ for the average iterate alongside $\tilde{\mathcal{O}}(T_{total}^{-1/3})$ for the last iterate in the Deterministic Regime. Here, $T_{total}$ denotes the total number of stochastic critic updates (or Monte Carlo rollouts). Furthermore, in the tabular setting, our positive-action-gap analysis yields a $\tilde{\mathcal{O}}(T_{total}^{-2/3})$ average-iterate rate, surpassing the $\mathcal{O}(T_{total}^{-1/2})$ worst-case statistical barrier that applies without a positive action margin.

cs.LG↗

arXiv2026

Bidirectional Resource Scheduling for Disaggregated and Asynchronous RL Post-Training

It is well established that the reasoning capabilities of large language models (LLMs) can be improved by applying reinforcement learning (RL) in a post-training stage. In a standard RL iteration, the current model (the policy) generates experience through rollouts, and the resulting data is then used to update the policy during training. High-performance RL frameworks such as StreamRL and AReaL employ a disaggregated architecture and asynchronous rollouts to better exploit both rollout and training resources, thereby increasing overall system throughput. Nonetheless, across varying RL setups (e.g., hardware configurations, model scales, staleness levels, and hyperparameters) and under changing workloads, it remains common for both rollout and training resources to experience idle periods. In this paper, we present BiDiRL, a hybrid time-space multiplexing architecture for asynchronous, disaggregated RL designed to reduce resource idleness. First, we develop a hot-switch runtime that enables rapid switching between rollout and training resources with negligible overhead. Second, we propose a static, scheduling-aware planner based on time-performance modeling that chooses a hot-switch-friendly resource partition, so that rollout and training durations are roughly balanced at a coarse level. Third, at execution time, we introduce a bidirectional scheduler that further exploits runtime bubbles through fine-grained resource switching, allowing the bottleneck stage to temporarily borrow idle resources from the other pool. Across a wide range of workloads, datasets, and models on two 32-GPU testbeds, BiDiRL increases RL training throughput by up to 1.94x compared with RL systems including veRL, AReaL, and ROLL, without affecting convergence behavior.

cs.DC↗

arXiv2026

Accelerating Disaggregated RL for Visual Generative LLMs with Diffusion-Based Parallelism and Trainer-Assisted Generation

Reinforcement learning (RL) has become a dominant post-training paradigm, driving the emergence of high-performance RL systems such as veRL for autoregressive large language models (LLMs). In parallel, diffusion-oriented RL algorithms, e.g., DanceGRPO and FlowGRPO, have rapidly expanded the scope of RL from language reasoning to diffusion-based visual and flow-based generation. However, efficient RL systems for diffusion generative LLMs remain underexplored. Existing implementations, e.g., veRL-Omni, still rely on colocated execution, which simplifies synchronization but couples rollout and training resources, limits heterogeneous deployment, and constrains independent scaling. To this end, we introduce DigenRL, a disaggregated RL framework for diffusion-based generative LLMs that supports flexible resource allocation, accommodates heterogeneous GPUs, and facilitates efficient task scheduling. To maximally reduce the execution bubbles in the disaggregated architecture, we propose: 1) a generation-axis pipeline (GAP) and time-step parallelism (TSP) in the diffusion architecture to enable finer-grained pipelining between rollout and training; 2) an elastic trainer-assisted generation (TAG) approach to enable the trainer GPU resources to dynamically assist in executing rollout generations; and 3) a tightly one-step constrained asynchronous strategy to further utilize the tail bubble in the pipeline. Extensive experiments are conducted on three hardware testbeds with 16-32 GPUs using HunyuanVideo-13B, Wan2.1-14B, FLUX.1-12B, and QwenImage-20B generative models. Experimental results show that DigenRL achieves 1.56-2.10x throughput improvements over state-of-the-art diffusion RL systems, veRL-Omni and GenRL.

cs.AI↗

arXiv2026

Re-examining and calibrating weighted survival analysis for causal inference

Causal inference with time-to-event outcomes is fundamental in various scientific studies. In a static setup with fitted propensity scores, weighted Kaplan-Meier estimation for survival probabilities and weighted Breslow-Peto estimation for hazard ratios have been widely used, but their statistical properties have been overlooked or studied only to a limited extent. We re-examine the weighted Kaplan-Meier method by formally linking it with the general framework of augmented inverse probability weighted estimation including both point and variance estimation. Furthermore, to address limitations of existing weighted methods for survival analysis, we develop new methods and associated theory through calibrated estimation in both low-dimensional and high-dimensional settings. We present a simulation study and an empirical application on the effectiveness of adjunctive psychotropic treatments for patients with schizophrenia. The calibrated methods yield coverage proportions closer to target ones in the simulation study, and produce shorter confidence intervals in both simulation and empirical studies.

stat.ME↗

arXiv2025

Preconditioned Discrete-HAMS: A Second-order Irreversible Discrete Sampler

Gradient-based Markov Chain Monte Carlo methods have recently received much attention for sampling discrete distributions, with notable examples such as Norm Constrained Gradient (NCG), Auxiliary Variable Gradient (AVG), and Discrete Hamiltonian Assisted Metropolis Sampling (DHAMS). In this work, we propose the Preconditioned Discrete-HAMS (PDHAMS) algorithm, which extends DHAMS by incorporating a second-order, quadratic approximation of the potential function, and uses Gaussian integral trick to avoid directly sampling a pairwise Markov random field. The PDHAMS sampler not only satisfies generalized detailed balance, hence enabling irreversible sampling, but also is a rejection-free property for a target distribution with a quadratic potential function. In various numerical experiments, PDHAMS algorithms consistently yield superior performance compared with other methods.

stat.ME↗

arXiv2025

Debiased Prediction Inference with Non-sparse Loadings in Misspecified High-dimensional Regression Models

High-dimensional regression models with regularized sparse estimation are widely applied. For statistical inferences, debiased methods are available about single coefficients or predictions with sparse new covariate vectors (also called loadings), in the presence of possible model misspecification. However, statistical inferences about predictions with non-sparse loadings are studied only under the assumption of correctly specified models. In this work, we develop debiased estimation and associated Wald confidence intervals for predictions with general loadings, allowed to be non-sparse, from possibly misspecified high-dimensional regression models. Our debiased estimator involves estimation of a debiasing vector, which is the general loading left-multiplied by the non-centered precision matrix in the linear model (LM) setting or the inverse Hessian of the objective function at the target coefficient vector in the generalized linear model (GLM) setting. We propose suitable estimators of the precision matrix or the inverse Hessian respectively in the LM or GLM settings and, for the first time, establish a root-n asymptotic expansion for the debiased prediction and justify associated Wald confidence intervals under sparsity conditions on the precision matrix or the inverse Hessian which are comparable to the conjunction of sparsity conditions required for inferences about all single coefficients in existing works. We also provide numerical results which further demonstrate the validity of our proposed confidence intervals for predictions with general loadings from possibly misspecified regression models.

math.ST↗

arXiv2025

Discrete Hamiltonian-Assisted Metropolis Sampling

Gradient-based Markov Chain Monte Carlo methods have recently received much attention for sampling discrete distributions, with interesting connections to their continuous counterparts. For examples, there are two discrete analogues to the Metropolis-adjusted Langevin Algorithm (MALA). As motivated by Hamiltonian-Assisted Metropolis Sampling (HAMS), we propose Discrete HAMS (DHAMS), a discrete sampler which, for the first time, not only exploits gradient information but also incorporates a Gaussian momentum variable and samples a Hamiltonian as an augmented distribution. DHAMS is derived through several steps, including an auxiliary-variable proposal scheme, negation and gradient correction for the momentum variable, and over-relaxation for the state variable. Two distinctive properties are achieved simultaneously. One is generalized detailed balance, which enables irreversible exploration of the target distribution. The other is a rejection-free property for a target distribution with a linear potential function. In experiments involving both ordinal and binary distributions, DHAMS algorithms consistently yield superior performance compared with existing algorithms.

stat.ME↗

arXiv2025

Distributionally enhanced marginal sensitivity model and bounds

For sensitivity analysis against unmeasured confounding, we build on the marginal sensitivity model (MSM) and propose a new model, deMSM, by incorporating a second constraint on the shift of potential outcome distributions caused by unmeasured confounders in addition to the constraint on the shift of treatment probabilities. We show that deMSM leads to interpretable sharp bounds of common causal parameters and tightens the corresponding MSM bounds. Moreover, the sharp bounds are symmetric in the two deMSM constraints, which facilitates practical applications. Lastly, we compare deMSM with other MSM-related models in both model constraints and sharp bounds, and reveal new interpretations for later models.

stat.ME↗

arXiv2025

Enhanced Marginal Sensitivity Model and Bounds

Sensitivity analysis is important to assess the impact of unmeasured confounding in causal inference from observational studies. The marginal sensitivity model (MSM) provides a useful approach in quantifying the influence of unmeasured confounders on treatment assignment and leading to tractable sharp bounds of common causal parameters. In this paper, to tighten MSM sharp bounds, we propose the enhanced MSM (eMSM) by incorporating another sensitivity constraint that quantifies the influence of unmeasured confounders on outcomes. We derive sharp population bounds of expected potential outcomes under eMSM, which are always narrower than the MSM sharp bounds in a simple and interpretable way. We further discuss desirable specifications of sensitivity parameters related to the outcome sensitivity constraint, and obtain both doubly robust point estimation and confidence intervals for the eMSM population bounds. The effectiveness of eMSM is also demonstrated numerically through two real-data applications. Our development represents, for the first time, a satisfactory extension of MSM to exploit both treatment and outcome sensitivity constraints on unmeasured confounding.

stat.ME↗

arXiv2024

Semi-supervised Regression Analysis with Model Misspecification and High-dimensional Data

The accessibility of vast volumes of unlabeled data has sparked growing interest in semi-supervised learning (SSL) and covariate shift transfer learning (CSTL). In this paper, we present an inference framework for estimating regression coefficients in conditional mean models within both SSL and CSTL settings, while allowing for the misspecification of conditional mean models. We develop an augmented inverse probability weighted (AIPW) method, employing regularized calibrated estimators for both propensity score (PS) and outcome regression (OR) nuisance models, with PS and OR models being sequentially dependent. We show that when the PS model is correctly specified, the proposed estimator achieves consistency, asymptotic normality, and valid confidence intervals, even with possible OR model misspecification and high-dimensional data. Moreover, by suppressing detailed technical choices, we demonstrate that previous methods can be unified within our AIPW framework. Our theoretical findings are verified through extensive simulation studies and a real-world data application.

stat.ME↗

arXiv2024

On Ridge Estimation in High-dimensional Rotationally Sparse Linear Regression

Recently, deep neural networks have been found to nearly interpolate training data but still generalize well in various applications. To help understand such a phenomenon, it has been of interest to analyze the ridge estimator and its interpolation limit in high-dimensional regression models. For this motivation, we study the ridge estimator in a rotationally sparse setting of high-dimensional linear regression, where the signal of a response is aligned with a small number, $d$, of covariates with large or spiked variances, compared with the remaining covariates with small or tail variances, \textit{after} an orthogonal transformation of the covariate vector. We establish high-probability upper and lower bounds on the out-sample and in-sample prediction errors in two distinct regimes depending on the ratio of the effective rank of tail variances over the sample size $n$. The separation of the two regimes enables us to exploit relevant concentration inequalities and derive concrete error bounds without making any oracle assumption or independent components assumption on covariate vectors. Moreover, we derive sufficient and necessary conditions which indicate that the prediction errors of ridge estimation can be of the order $O(\frac{d}{n})$ if and only if the gap between the spiked and tail variances are sufficiently large. We also compare the orders of optimal out-sample and in-sample prediction errors and find that, remarkably, the optimal out-sample prediction error may be significantly smaller than the optimal in-sample one. Finally, we present numerical experiments which empirically confirm our theoretical findings.

math.ST↗

arXiv2023

On semi-supervised estimation using exponential tilt mixture models

Consider a semi-supervised setting with a labeled dataset of binary responses and predictors and an unlabeled dataset with only the predictors. Logistic regression is equivalent to an exponential tilt model in the labeled population. For semi-supervised estimation, we develop further analysis and understanding of a statistical approach using exponential tilt mixture (ETM) models and maximum nonparametric likelihood estimation, while allowing that the class proportions may differ between the unlabeled and labeled data. We derive asymptotic properties of ETM-based estimation and demonstrate improved efficiency over supervised logistic regression in a random sampling setup and an outcome-stratified sampling setup previously used. Moreover, we reconcile such efficiency improvement with the existing semiparametric efficiency theory when the class proportions in the unlabeled and labeled data are restricted to be the same. We also provide a simulation study to numerically illustrate our theoretical findings.

stat.ML↗

arXiv2023

Sensitivity models and bounds under sequential unmeasured confounding in longitudinal studies

Consider sensitivity analysis to assess the worst-case possible values of counterfactual outcome means and average treatment effects under sequential unmeasured confounding in a longitudinal study with time-varying treatments and covariates. We formulate several multi-period sensitivity models to relax the corresponding versions of the assumption of sequential unconfounding. The primary sensitivity model involves only counterfactual outcomes, whereas the joint and product sensitivity models involve both counterfactual covariates and outcomes. We establish and compare explicit representations for the sharp and conservative bounds at the population level through convex optimization, depending only on the observed data. These results provide for the first time a satisfactory generalization from the marginal sensitivity model in the cross-sectional setting.

math.ST↗

arXiv2023

Persistently Trained, Diffusion-assisted Energy-based Models

Maximum likelihood (ML) learning for energy-based models (EBMs) is challenging, partly due to non-convergence of Markov chain Monte Carlo.Several variations of ML learning have been proposed, but existing methods all fail to achieve both post-training image generation and proper density estimation. We propose to introduce diffusion data and learn a joint EBM, called diffusion assisted-EBMs, through persistent training (i.e., using persistent contrastive divergence) with an enhanced sampling algorithm to properly sample from complex, multimodal distributions. We present results from a 2D illustrative experiment and image experiments and demonstrate that, for the first time for image data, persistently trained EBMs can {\it simultaneously} achieve long-run stability, post-training image generation, and superior out-of-distribution detection.

stat.ML↗

arXiv2023

Understanding Accelerated Gradient Methods: Lyapunov Analyses and Hamiltonian Assisted Interpretations

We formulate two classes of first-order algorithms more general than previously studied for minimizing smooth and strongly convex or, respectively, smooth and convex functions. We establish sufficient conditions, via new discrete Lyapunov analyses, for achieving accelerated convergence rates which match Nesterov's methods in the strongly and general convex settings. Next, we study the convergence of limiting ordinary differential equations (ODEs) and point out currently notable gaps between the convergence properties of the corresponding algorithms and ODEs. Finally, we propose a novel class of discrete algorithms, called the Hamiltonian assisted gradient method, directly based on a Hamiltonian function and several interpretable operations, and then demonstrate meaningful and unified interpretations of our acceleration conditions.

math.OC↗

arXiv2022

Block-wise Primal-dual Algorithms for Large-scale Doubly Penalized ANOVA Modeling

For multivariate nonparametric regression, doubly penalized ANOVA modeling (DPAM) has recently been proposed, using hierarchical total variations (HTVs) and empirical norms as penalties on the component functions such as main effects and multi-way interactions in a functional ANOVA decomposition of the underlying regression function. The two penalties play complementary roles: the HTV penalty promotes sparsity in the selection of basis functions within each component function, whereas the empirical-norm penalty promotes sparsity in the selection of component functions. We adopt backfitting or block minimization for training DPAM, and develop two suitable primal-dual algorithms, including both batch and stochastic versions, for updating each component function in single-block optimization. Existing applications of primal-dual algorithms are intractable in our setting with both HTV and empirical-norm penalties. Through extensive numerical experiments, we demonstrate the validity and advantage of our stochastic primal-dual algorithms, compared with their batch versions and a previous active-set algorithm, in large-scale scenarios.

stat.CO↗

arXiv2022

Model-assisted sensitivity analysis for treatment effects under unmeasured confounding via regularized calibrated estimation

Consider sensitivity analysis for estimating average treatment effects under unmeasured confounding, assumed to satisfy a marginal sensitivity model. At the population level, we provide new representations for the sharp population bounds and doubly robust estimating functions, recently derived by Dorn, Guo, and Kallus. We also derive new, relaxed population bounds, depending on weighted linear outcome quantile regression. At the sample level, we develop new methods and theory for obtaining not only doubly robust point estimators for the relaxed population bounds with respect to misspecification of a propensity score model or an outcome mean regression model, but also model-assisted confidence intervals which are valid if the propensity score model is correctly specified, but the outcome quantile and mean regression models may be misspecified. The relaxed population bounds reduce to the sharp bounds if outcome quantile regression is correctly specified. For a linear outcome mean regression model, the confidence intervals are also doubly robust. Our methods involve regularized calibrated estimation, with Lasso penalties but carefully chosen loss functions, for fitting propensity score and outcome mean and quantile regression models. We present a simulation study and an empirical application to an observational study on the effects of right heart catheterization.

stat.ME↗