Qi Feng

Wei Deng, Yu Chen, Nicole Tianjiao Yang et al.

Diffusion models have become the go-to method for large-scale generative models in real-world applications. These applications often involve data distributions confined within bounded domains, typically requiring ad-hoc thresholding techniques for boundary enforcement. Reflected diffusion models (Lou23) aim to enhance generalizability by generating the data distribution through a backward process governed by reflected Brownian motion. However, reflected diffusion models may not easily adapt to diverse domains without the derivation of proper diffeomorphic mappings and do not guarantee optimal transport properties. To overcome these limitations, we introduce the Reflected Schrodinger Bridge algorithm: an entropy-regularized optimal transport approach tailored for generating data within diverse bounded domains. We derive elegant reflected forward-backward stochastic differential equations with Neumann and Robin boundary conditions, extend divergence-based likelihood training to bounded domains, and explore natural connections to entropic optimal transport for the study of approximate linear convergence - a valuable insight for practical training. Our algorithm yields robust generative modeling in diverse domains, and its scalability is demonstrated in real-world constrained generative modeling through standard image benchmarks.

Qi Feng, Xinzhe Zuo, Wuchen Li

We provide a Lyapunov convergence analysis for time-inhomogeneous variable coefficient stochastic differential equations (SDEs). Three typical examples include overdamped, irreversible drift, and underdamped Langevin dynamics. We first formula the probability transition equation of Langevin dynamics as a modified gradient flow of the Kullback-Leibler divergence in the probability space with respect to time-dependent optimal transport metrics. This formulation contains both gradient and non-gradient directions depending on a class of time-dependent target distribution. We then select a time-dependent relative Fisher information functional as a Lyapunov functional. We develop a time-dependent Hessian matrix condition, which guarantees the convergence of the probability density function of the SDE. We verify the proposed conditions for several time-inhomogeneous Langevin dynamics. For the overdamped Langevin dynamics, we prove the $O(t^{-1/2})$ convergence in $L^1$ distance for the simulated annealing dynamics with a strongly convex potential function. For the irreversible drift Langevin dynamics, we prove an improved convergence towards the target distribution in an asymptotic regime. We also verify the convergence condition for the underdamped Langevin dynamics. Numerical examples demonstrate the convergence results for the time-dependent Langevin dynamics.

Juncheng Dong, Yang Yang, Tao Liu et al.

The efficiency of GPU kernels is central to the progress of modern AI, yet optimizing them remains a difficult and labor-intensive task due to complex interactions between memory hierarchies, thread scheduling, and hardware-specific characteristics. While recent advances in large language models (LLMs) provide new opportunities for automated code generation, existing approaches largely treat LLMs as single-shot generators or naive refinement tools, limiting their effectiveness in navigating the irregular kernel optimization landscape. We introduce an LLM agentic framework for GPU kernel optimization that systematically explores the design space through multi-agent collaboration, grounded instruction, dynamic context management, and strategic search. This framework mimics the workflow of expert engineers, enabling LLMs to reason about hardware trade-offs, incorporate profiling feedback, and refine kernels iteratively. We evaluate our approach on KernelBench, a benchmark for LLM-based kernel optimization, and demonstrate substantial improvements over baseline agents: our system produces correct solutions where baselines often fail, and achieves kernels with up to 16x faster runtime performance. These results highlight the potential of agentic LLM frameworks to advance fully automated, scalable GPU kernel optimization.

Yitao Shen, Yue Wang, Xingyu Lu et al.

Technology companies building consumer-facing platforms may have access to massive-scale user population. In recent years, promotion with quantifiable incentive has become a popular approach for increasing active users on such platforms. On one hand, increased user activities can introduce network effect, bring in advertisement audience, and produce other benefits. On the other hand, massive-scale promotion causes massive cost. Therefore making promotion campaigns efficient in terms of return-on-investment (ROI) is of great interest to many companies. This paper proposes a practical two-stage framework that can optimize the ROI of various massive-scale promotion campaigns. In the first stage, users' personal promotion-response curves are modeled by machine learning techniques. In the second stage, business objectives and resource constraints are formulated into an optimization problem, the decision variables of which are how much incentive to give to each user. In order to do effective optimization in the second stage, counterfactual prediction and noise-reduction are essential for the first stage. We leverage existing counterfactual prediction techniques to correct treatment bias in data. We also introduce a novel deep neural network (DNN) architecture, the deep-isotonic-promotion-network (DIPN), to reduce noise in the promotion response curves. The DIPN architecture incorporates our prior knowledge of response curve shape, by enforcing isotonicity and smoothness. It out-performed regular DNN and other state-of-the-art shape-constrained models in our experiments.

Runzhong Wang, Zhigang Hua, Gan Liu et al.

Combinatorial Optimization (CO) has been a long-standing challenging research topic featured by its NP-hard nature. Traditionally such problems are approximately solved with heuristic algorithms which are usually fast but may sacrifice the solution quality. Currently, machine learning for combinatorial optimization (MLCO) has become a trending research topic, but most existing MLCO methods treat CO as a single-level optimization by directly learning the end-to-end solutions, which are hard to scale up and mostly limited by the capacity of ML models given the high complexity of CO. In this paper, we propose a hybrid approach to combine the best of the two worlds, in which a bi-level framework is developed with an upper-level learning method to optimize the graph (e.g. add, delete or modify edges in a graph), fused with a lower-level heuristic algorithm solving on the optimized graph. Such a bi-level approach simplifies the learning on the original hard CO and can effectively mitigate the demand for model capacity. The experiments and results on several popular CO problems like Directed Acyclic Graph scheduling, Graph Edit Distance and Hamiltonian Cycle Problem show its effectiveness over manually designed heuristics and single-level learning methods.

Zhigang Hua, Feng Qi, Gan Liu et al.

Scheduling computational tasks represented by directed acyclic graphs (DAGs) is challenging because of its complexity. Conventional scheduling algorithms rely heavily on simple heuristics such as shortest job first (SJF) and critical path (CP), and are often lacking in scheduling quality. In this paper, we present a novel learning-based approach to scheduling DAG tasks. The algorithm employs a reinforcement learning agent to iteratively add directed edges to the DAG, one at a time, to enforce ordering (i.e., priorities of execution and resource allocation) of "tricky" job nodes. By doing so, the original DAG scheduling problem is dramatically reduced to a much simpler proxy problem, on which heuristic scheduling algorithms such as SJF and CP can be efficiently improved. Our approach can be easily applied to any existing heuristic scheduling algorithms. On the benchmark dataset of TPC-H, we show that our learning based approach can significantly improve over popular heuristic algorithms and consistently achieves the best performance among several methods under a variety of settings.

Wei Deng, Qi Feng, Georgios Karagiannis et al.

Replica exchange stochastic gradient Langevin dynamics (reSGLD) has shown promise in accelerating the convergence in non-convex learning; however, an excessively large correction for avoiding biases from noisy energy estimators has limited the potential of the acceleration. To address this issue, we study the variance reduction for noisy energy estimators, which promotes much more effective swaps. Theoretically, we provide a non-asymptotic analysis on the exponential acceleration for the underlying continuous-time Markov jump process; moreover, we consider a generalized Girsanov theorem which includes the change of Poisson measure to overcome the crude discretization based on the Gröwall's inequality and yields a much tighter error in the 2-Wasserstein ($\mathcal{W}_2$) distance. Numerically, we conduct extensive experiments and obtain the state-of-the-art results in optimization and uncertainty estimates for synthetic experiments and image data.

Wei Deng, Qi Feng, Liyao Gao et al.

Replica exchange Monte Carlo (reMC), also known as parallel tempering, is an important technique for accelerating the convergence of the conventional Markov Chain Monte Carlo (MCMC) algorithms. However, such a method requires the evaluation of the energy function based on the full dataset and is not scalable to big data. The naïve implementation of reMC in mini-batch settings introduces large biases, which cannot be directly extended to the stochastic gradient MCMC (SGMCMC), the standard sampling method for simulating from deep neural networks (DNNs). In this paper, we propose an adaptive replica exchange SGMCMC (reSGMCMC) to automatically correct the bias and study the corresponding properties. The analysis implies an acceleration-accuracy trade-off in the numerical discretization of a Markov jump process in a stochastic environment. Empirically, we test the algorithm through extensive experiments on various setups and obtain the state-of-the-art results on CIFAR10, CIFAR100, and SVHN in both supervised learning and semi-supervised learning tasks.