arXiv subjects

Xiaocheng Tang

Publications and source records attributed to Xiaocheng Tang.

17 recordsLinked to original sources

arXiv2026

Beyond Reasoning Gains: Mitigating General-Capability Forgetting in Large Reasoning Models

Reinforcement learning with verifiable rewards (RLVR) has delivered impressive gains in mathematical and multimodal reasoning and has become a standard post-training paradigm for contemporary language and vision-language models. However, the RLVR recipe introduces a significant risk of capability regression, in which models forget foundational skills after prolonged training without employing regularization strategies. We empirically confirm this concern, observing that open-source reasoning models suffer performance degradation on core capabilities such as perception and faithfulness. While imposing regularization terms like KL divergence can help prevent deviation from the base model, these terms are computed on the current task and therefore do not guarantee preservation of broader knowledge. Meanwhile, commonly used experience replay across heterogeneous domains makes it nontrivial to decide how much training emphasis each objective should receive. To address this, we propose RECAP-a replay strategy with dynamic objective reweighting for general knowledge preservation. Our reweighting mechanism adapts online using short-horizon signals of convergence and instability, shifting the post-training focus away from saturated objectives and toward underperforming or volatile ones. Our method is end-to-end and readily applicable to existing RLVR pipelines without training additional models or heavy tuning. Extensive experiments on benchmarks using Qwen2.5-VL-3B and Qwen2.5-VL-7B demonstrate the effectiveness of our method, which not only preserves general capabilities but also improves reasoning by enabling more flexible trade-offs among in-task rewards.

cs.LG↗

arXiv2025

LlamaRL: A Distributed Asynchronous Reinforcement Learning Framework for Efficient Large-scale LLM Training

Reinforcement Learning (RL) has become the most effective post-training approach for improving the capabilities of Large Language Models (LLMs). In practice, because of the high demands on latency and memory, it is particularly challenging to develop an efficient RL framework that reliably manages policy models with hundreds to thousands of billions of parameters. In this paper, we present LlamaRL, a fully distributed, asynchronous RL framework optimized for efficient training of large-scale LLMs with various model sizes (8B, 70B, and 405B parameters) on GPU clusters ranging from a handful to thousands of devices. LlamaRL introduces a streamlined, single-controller architecture built entirely on native PyTorch, enabling modularity, ease of use, and seamless scalability to thousands of GPUs. We also provide a theoretical analysis of LlamaRL's efficiency, including a formal proof that its asynchronous design leads to strict RL speed-up. Empirically during the Llama 3 post-training, by leveraging best practices such as colocated model offloading, asynchronous off-policy training, and distributed direct memory access for weight synchronization, LlamaRL achieves significant efficiency gains -- up to 10.7x speed-up compared to DeepSpeed-Chat-like systems on a 405B-parameter policy model. Furthermore, the efficiency advantage continues to grow with increasing model scale, demonstrating the framework's suitability for future large-scale RL training.

cs.LG↗

arXiv2025

Spatio-temporal Prediction of Fine-Grained Origin-Destination Matrices with Applications in Ridesharing

Accurate spatial-temporal prediction of network-based travelers' requests is crucial for the effective policy design of ridesharing platforms. Having knowledge of the total demand between various locations in the upcoming time slots enables platforms to proactively prepare adequate supplies, thereby increasing the likelihood of fulfilling travelers' requests and redistributing idle drivers to areas with high potential demand to optimize the global supply-demand equilibrium. This paper delves into the prediction of Origin-Destination (OD) demands at a fine-grained spatial level, especially when confronted with an expansive set of local regions. While this task holds immense practical value, it remains relatively unexplored within the research community. To fill this gap, we introduce a novel prediction model called OD-CED, which comprises an unsupervised space coarsening technique to alleviate data sparsity and an encoder-decoder architecture to capture both semantic and geographic dependencies. Through practical experimentation, OD-CED has demonstrated remarkable results. It achieved an impressive reduction of up to 45% reduction in root-mean-square error and 60% in weighted mean absolute percentage error over traditional statistical methods when dealing with OD matrices exhibiting a sparsity exceeding 90%.

cs.LG↗

arXiv2022

Golfer: Trajectory Prediction with Masked Goal Conditioning MnM Network

Transformers have enabled breakthroughs in NLP and computer vision, and have recently began to show promising performance in trajectory prediction for Autonomous Vehicle (AV). How to efficiently model the interactive relationships between the ego agent and other road and dynamic objects remains challenging for the standard attention module. In this work we propose a general Transformer-like architectural module MnM network equipped with novel masked goal conditioning training procedures for AV trajectory prediction. The resulted model, named golfer, achieves state-of-the-art performance, winning the 2nd place in the 2022 Waymo Open Dataset Motion Prediction Challenge and ranked 1st place according to minADE.

cs.CV↗

arXiv2022

Reinforcement Learning in the Wild: Scalable RL Dispatching Algorithm Deployed in Ridehailing Marketplace

In this study, a real-time dispatching algorithm based on reinforcement learning is proposed and for the first time, is deployed in large scale. Current dispatching methods in ridehailing platforms are dominantly based on myopic or rule-based non-myopic approaches. Reinforcement learning enables dispatching policies that are informed of historical data and able to employ the learned information to optimize returns of expected future trajectories. Previous studies in this field yielded promising results, yet have left room for further improvements in terms of performance gain, self-dependency, transferability, and scalable deployment mechanisms. The present study proposes a standalone RL-based dispatching solution that is equipped with multiple mechanisms to ensure robust and efficient on-policy learning and inference while being adaptable for full-scale deployment. A new form of value updating based on temporal difference is proposed that is more adapted to the inherent uncertainty of the problem. For the driver-order assignment, a customized utility function is proposed that when tuned based on the statistics of the market, results in remarkable performance improvement and interpretability. In addition, for reducing the risk of cancellation after drivers' assignment, an adaptive graph pruning strategy based on the multi-arm bandit problem is introduced. The method is evaluated using offline simulation with real data and yields notable performance improvement. In addition, the algorithm is deployed online in multiple cities under DiDi's operation for A/B testing and is launched in one of the major international markets as the primary mode of dispatch. The deployed algorithm shows over 1.3% improvement in total driver income from A/B testing. In addition, by causal inference analysis, as much as 5.3% improvement in major performance metrics is detected after full-scale deployment.

cs.LG↗

arXiv2021

RANK-NOSH: Efficient Predictor-Based Architecture Search via Non-Uniform Successive Halving

Predictor-based algorithms have achieved remarkable performance in the Neural Architecture Search (NAS) tasks. However, these methods suffer from high computation costs, as training the performance predictor usually requires training and evaluating hundreds of architectures from scratch. Previous works along this line mainly focus on reducing the number of architectures required to fit the predictor. In this work, we tackle this challenge from a different perspective - improve search efficiency by cutting down the computation budget of architecture training. We propose NOn-uniform Successive Halving (NOSH), a hierarchical scheduling algorithm that terminates the training of underperforming architectures early to avoid wasting budget. To effectively leverage the non-uniform supervision signals produced by NOSH, we formulate predictor-based architecture search as learning to rank with pairwise comparisons. The resulting method - RANK-NOSH, reduces the search budget by ~5x while achieving competitive or even better performance than previous state-of-the-art predictor-based methods on various spaces and datasets.

cs.LG↗

arXiv2021

Rethinking Architecture Selection in Differentiable NAS

Differentiable Neural Architecture Search is one of the most popular Neural Architecture Search (NAS) methods for its search efficiency and simplicity, accomplished by jointly optimizing the model weight and architecture parameters in a weight-sharing supernet via gradient-based algorithms. At the end of the search phase, the operations with the largest architecture parameters will be selected to form the final architecture, with the implicit assumption that the values of architecture parameters reflect the operation strength. While much has been discussed about the supernet's optimization, the architecture selection process has received little attention. We provide empirical and theoretical analysis to show that the magnitude of architecture parameters does not necessarily indicate how much the operation contributes to the supernet's performance. We propose an alternative perturbation-based architecture selection that directly measures each operation's influence on the supernet. We re-evaluate several differentiable NAS methods with the proposed architecture selection and find that it is able to extract significantly improved architectures from the underlying supernets consistently. Furthermore, we find that several failure modes of DARTS can be greatly alleviated with the proposed selection method, indicating that much of the poor generalization observed in DARTS can be attributed to the failure of magnitude-based architecture selection rather than entirely the optimization of its supernet.

cs.LG↗

arXiv2021

Real-world Ride-hailing Vehicle Repositioning using Deep Reinforcement Learning

We present a new practical framework based on deep reinforcement learning and decision-time planning for real-world vehicle repositioning on ride-hailing (a type of mobility-on-demand, MoD) platforms. Our approach learns the spatiotemporal state-value function using a batch training algorithm with deep value networks. The optimal repositioning action is generated on-demand through value-based policy search, which combines planning and bootstrapping with the value networks. For the large-fleet problems, we develop several algorithmic features that we incorporate into our framework and that we demonstrate to induce coordination among the algorithmically-guided vehicles. We benchmark our algorithm with baselines in a ride-hailing simulation environment to demonstrate its superiority in improving income efficiency meausred by income-per-hour. We have also designed and run a real-world experiment program with regular drivers on a major ride-hailing platform. We have observed significantly positive results on key metrics comparing our method with experienced drivers who performed idle-time repositioning based on their own expertise.

cs.LG↗

arXiv2021

A Deep Value-network Based Approach for Multi-Driver Order Dispatching

Recent works on ride-sharing order dispatching have highlighted the importance of taking into account both the spatial and temporal dynamics in the dispatching process for improving the transportation system efficiency. At the same time, deep reinforcement learning has advanced to the point where it achieves superhuman performance in a number of fields. In this work, we propose a deep reinforcement learning based solution for order dispatching and we conduct large scale online A/B tests on DiDi's ride-dispatching platform to show that the proposed method achieves significant improvement on both total driver income and user experience related metrics. In particular, we model the ride dispatching problem as a Semi Markov Decision Process to account for the temporal aspect of the dispatching actions. To improve the stability of the value iteration with nonlinear function approximators like neural networks, we propose Cerebellar Value Networks (CVNet) with a novel distributed state representation layer. We further derive a regularized policy evaluation scheme for CVNet that penalizes large Lipschitz constant of the value network for additional robustness against adversarial perturbation and noises. Finally, we adapt various transfer learning methods to CVNet for increased learning adaptability and efficiency across multiple cities. We conduct extensive offline simulations based on real dispatching data as well as online AB tests through the DiDi's platform. Results show that CVNet consistently outperforms other recently proposed dispatching methods. We finally show that the performance can be further improved through the efficient use of transfer learning.

cs.LG↗

arXiv2021

Value Function is All You Need: A Unified Learning Framework for Ride Hailing Platforms

Large ride-hailing platforms, such as DiDi, Uber and Lyft, connect tens of thousands of vehicles in a city to millions of ride demands throughout the day, providing great promises for improving transportation efficiency through the tasks of order dispatching and vehicle repositioning. Existing studies, however, usually consider the two tasks in simplified settings that hardly address the complex interactions between the two, the real-time fluctuations between supply and demand, and the necessary coordinations due to the large-scale nature of the problem. In this paper we propose a unified value-based dynamic learning framework (V1D3) for tackling both tasks. At the center of the framework is a globally shared value function that is updated continuously using online experiences generated from real-time platform transactions. To improve the sample-efficiency and the robustness, we further propose a novel periodic ensemble method combining the fast online learning with a large-scale offline training scheme that leverages the abundant historical driver trajectory data. This allows the proposed framework to adapt quickly to the highly dynamic environment, to generalize robustly to recurrent patterns and to drive implicit coordinations among the population of managed vehicles. Extensive experiments based on real-world datasets show considerably improvements over other recently proposed methods on both tasks. Particularly, V1D3 outperforms the first prize winners of both dispatching and repositioning tracks in the KDD Cup 2020 RL competition, achieving state-of-the-art results on improving both total driver income and user experience related metrics.

cs.LG↗

arXiv2021

Measuring Sample Efficiency and Generalization in Reinforcement Learning Benchmarks: NeurIPS 2020 Procgen Benchmark

The NeurIPS 2020 Procgen Competition was designed as a centralized benchmark with clearly defined tasks for measuring Sample Efficiency and Generalization in Reinforcement Learning. Generalization remains one of the most fundamental challenges in deep reinforcement learning, and yet we do not have enough benchmarks to measure the progress of the community on Generalization in Reinforcement Learning. We present the design of a centralized benchmark for Reinforcement Learning which can help measure Sample Efficiency and Generalization in Reinforcement Learning by doing end to end evaluation of the training and rollout phases of thousands of user submitted code bases in a scalable way. We designed the benchmark on top of the already existing Procgen Benchmark by defining clear tasks and standardizing the end to end evaluation setups. The design aims to maximize the flexibility available for researchers who wish to design future iterations of such benchmarks, and yet imposes necessary practical constraints to allow for a system like this to scale. This paper presents the competition setup and the details and analysis of the top solutions identified through this setup in context of 2020 iteration of the competition at NeurIPS.

cs.LG↗

arXiv2021

DrNAS: Dirichlet Neural Architecture Search

This paper proposes a novel differentiable architecture search method by formulating it into a distribution learning problem. We treat the continuously relaxed architecture mixing weight as random variables, modeled by Dirichlet distribution. With recently developed pathwise derivatives, the Dirichlet parameters can be easily optimized with gradient-based optimizer in an end-to-end manner. This formulation improves the generalization ability and induces stochasticity that naturally encourages exploration in the search space. Furthermore, to alleviate the large memory consumption of differentiable NAS, we propose a simple yet effective progressive learning scheme that enables searching directly on large-scale tasks, eliminating the gap between search and evaluation phases. Extensive experiments demonstrate the effectiveness of our method. Specifically, we obtain a test error of 2.46% for CIFAR-10, 23.7% for ImageNet under the mobile setting. On NAS-Bench-201, we also achieve state-of-the-art results on all three datasets and provide insights for the effective design of neural architecture search algorithms.

cs.LG↗

arXiv2019

Deep Reinforcement Learning for Multi-Driver Vehicle Dispatching and Repositioning Problem

Order dispatching and driver repositioning (also known as fleet management) in the face of spatially and temporally varying supply and demand are central to a ride-sharing platform marketplace. Hand-crafting heuristic solutions that account for the dynamics in these resource allocation problems is difficult, and may be better handled by an end-to-end machine learning method. Previous works have explored machine learning methods to the problem from a high-level perspective, where the learning method is responsible for either repositioning the drivers or dispatching orders, and as a further simplification, the drivers are considered independent agents maximizing their own reward functions. In this paper we present a deep reinforcement learning approach for tackling the full fleet management and dispatching problems. In addition to treating the drivers as individual agents, we consider the problem from a system-centric perspective, where a central fleet management agent is responsible for decision-making for all drivers.

cs.LG↗

arXiv2019

CoRide: Joint Order Dispatching and Fleet Management for Multi-Scale Ride-Hailing Platforms

How to optimally dispatch orders to vehicles and how to tradeoff between immediate and future returns are fundamental questions for a typical ride-hailing platform. We model ride-hailing as a large-scale parallel ranking problem and study the joint decision-making task of order dispatching and fleet management in online ride-hailing platforms. This task brings unique challenges in the following four aspects. First, to facilitate a huge number of vehicles to act and learn efficiently and robustly, we treat each region cell as an agent and build a multi-agent reinforcement learning framework. Second, to coordinate the agents from different regions to achieve long-term benefits, we leverage the geographical hierarchy of the region grids to perform hierarchical reinforcement learning. Third, to deal with the heterogeneous and variant action space for joint order dispatching and fleet management, we design the action as the ranking weight vector to rank and select the specific order or the fleet management destination in a unified formulation. Fourth, to achieve the multi-scale ride-hailing platform, we conduct the decision-making process in a hierarchical way where a multi-head attention mechanism is utilized to incorporate the impacts of neighbor agents and capture the key agent in each scale. The whole novel framework is named as CoRide. Extensive experiments based on multiple cities real-world data as well as analytic synthetic data demonstrate that CoRide provides superior performance in terms of platform revenue and user experience in the task of city-wide hybrid order dispatching and fleet management over strong baselines.

cs.MA↗

arXiv2015

Practical Inexact Proximal Quasi-Newton Method with Global Complexity Analysis

Recently several methods were proposed for sparse optimization which make careful use of second-order information [10, 28, 16, 3] to improve local convergence rates. These methods construct a composite quadratic approximation using Hessian information, optimize this approximation using a first-order method, such as coordinate descent and employ a line search to ensure sufficient descent. Here we propose a general framework, which includes slightly modified versions of existing algorithms and also a new algorithm, which uses limited memory BFGS Hessian approximations, and provide a novel global convergence rate analysis, which covers methods that solve subproblems via coordinate descent.

cs.LG↗

arXiv2014

HIPAD - A Hybrid Interior-Point Alternating Direction algorithm for knowledge-based SVM and feature selection

We consider classification tasks in the regime of scarce labeled training data in high dimensional feature space, where specific expert knowledge is also available. We propose a new hybrid optimization algorithm that solves the elastic-net support vector machine (SVM) through an alternating direction method of multipliers in the first phase, followed by an interior-point method for the classical SVM in the second phase. Both SVM formulations are adapted to knowledge incorporation. Our proposed algorithm addresses the challenges of automatic feature selection, high optimization accuracy, and algorithmic flexibility for taking advantage of prior knowledge. We demonstrate the effectiveness and efficiency of our algorithm and compare it with existing methods on a collection of synthetic and real-world data.

stat.ML↗

arXiv2013

Efficiently Using Second Order Information in Large l1 Regularization Problems

We propose a novel general algorithm LHAC that efficiently uses second-order information to train a class of large-scale l1-regularized problems. Our method executes cheap iterations while achieving fast local convergence rate by exploiting the special structure of a low-rank matrix, constructed via quasi-Newton approximation of the Hessian of the smooth loss function. A greedy active-set strategy, based on the largest violations in the dual constraints, is employed to maintain a working set that iteratively estimates the complement of the optimal active set. This allows for smaller size of subproblems and eventually identifies the optimal active set. Empirical comparisons confirm that LHAC is highly competitive with several recently proposed state-of-the-art specialized solvers for sparse logistic regression and sparse inverse covariance matrix selection.

stat.ML↗