Guangwen Yang

Weitao Li, Hao Zhou, Xuanyu Lei et al.

Reinforcement learning with verifiable rewards (RLVR) has become central to LLM reasoning, but its outcome-level rewards can make models more willing to give confident answers when evidence or reasoning is unreliable. Existing SFT or RL methods mainly teach LLMs to refuse or express uncertainty at the response level, which can overfit abstention behavior rather than improve reasoning reliability. To address this limitation, we propose Cognitive Pairwise Training (CPT), a cognitive mid-training alignment stage that turns pairwise comparisons over reasoning traces into a reusable alignment signal. By learning to distinguish trustworthy from flawed reasoning, CPT encourages the model to internalize a reasoning-quality discrimination boundary rather than memorize surface refusal patterns. Across five model scales and three model families, CPT improves the reasoning--metacognition trade-off. At 14B, CPT+RL outperforms the standard SFT+RL pipeline by +2.2 math-average points and +5.2 abstention-F1 points. Further analyses show that CPT improves trace quality and exhibits strong robustness and scalability across evaluation and training settings. Code and models are released at https://github.com/Tsinghua-dhy/CPT.

Xin Wang, Wei Xue, Yilun Han et al.

Hybrid modeling combining data-driven techniques and numerical methods is an emerging and promising research direction for efficient climate simulation. However, previous works lack practical platforms, making developing hybrid modeling a challenging programming problem. Furthermore, the lack of standard data sets and evaluation metrics may hamper researchers from comprehensively comparing various algorithms under a uniform condition. To address these problems, we propose a framework called NeuroClim for hybrid modeling under the real-world scenario, a basic setting to simulate the real climate that we live in. NeuroClim consists of three parts: (1) Platform. We develop a user-friendly platform NeuroGCM for efficiently developing hybrid modeling in climate simulation. (2) Dataset. We provide an open-source dataset for data-driven methods in hybrid modeling. We investigate the characteristics of the data, i.e., heterogeneity and stiffness, which reveals the difficulty of regressing climate simulation data; (3) Metrics. We propose a methodology for quantitatively evaluating hybrid modeling, including the approximation ability of machine learning models and the stability during simulation. We believe that NeuroClim allows researchers to work without high level of climate-related expertise and focus only on machine learning algorithm design, which will accelerate hybrid modeling research in the AI-Climate intersection. The codes and data are released at https://github.com/x-w19/NeuroClim.

Yufan Jiang, Qiaozhi He, Xiaomin Zhuang et al. · tsinghua

Existing large language models have to run K times to generate a sequence of K tokens. In this paper, we present RecycleGPT, a generative language model with fast decoding speed by recycling pre-generated model states without running the whole model in multiple steps. Our approach relies on the observation that adjacent tokens in a sequence usually have strong correlations and the next token in a sequence can be reasonably guessed or inferred based on the preceding ones. Experiments and analysis demonstrate the effectiveness of our approach in lowering inference latency, achieving up to 1.4x speedup while preserving high performance.

Yuchuan Tian, Yuchen Liang, Jiacheng Sun et al.

Large language models (LLMs) excel at generation but dominant autoregressive (AR) decoding is inherently sequential, creating a throughput bottleneck. Diffusion Language Models (DLMs)--especially block-wise variants--enable parallel generation and intra-block bidirectional reasoning, yet training large DLMs from scratch is costly and wastes the knowledge in mature AR checkpoints. Prior "adaptation" attempts either modify logits or randomly grow attention masks to full-sequence diffusion, or simply transplant AR weights into a block-diffusion recipe, leaving a fundamental mismatch between AR causality and block-wise bidirectionality unaddressed. We reframe adaptation as a intra-paradigm path from AR to Block-Diffusion by viewing AR as Block-Diffusion with blocksize=1. Concretely, we design the pathway of adaptation as follows: we use a context-causal attention mask (causal in context, bidirectional only within the active block), an efficient parallel adaptation procedure, an auxiliary AR loss to maximize data utilization and retain pretrained knowledge, and gradual increment of the generation block size. The recipe integrates cleanly with masked block-diffusion and maintains train-inference consistency. Built on these components, NBDiff-7B (Base and Instruct) could inherit the long-context modeling and reasoning capabilities, and achieve state-of-the-art performance among the 7B-class DLMs, delivering strong gains on general-knowledge, math, and code benchmarks over strong baselines. These results demonstrate that principled AR-to-block-diffusion adaptation is an effective and compute-efficient alternative to training DLMs from scratch. Codes: https://github.com/YuchuanTian/NBDiff.

Jianing Ye, Chenghao Li, Yongqiang Dou et al.

Decentralized execution is one core demand in multi-agent reinforcement learning (MARL). Recently, most popular MARL algorithms have adopted decentralized policies to enable decentralized execution, and use gradient descent as the optimizer. However, there is hardly any theoretical analysis of these algorithms taking the optimization method into consideration, and we find that various popular MARL algorithms with decentralized policies are suboptimal in toy tasks when gradient descent is chosen as their optimization method. In this paper, we theoretically analyze two common classes of algorithms with decentralized policies -- multi-agent policy gradient methods and value-decomposition methods, and prove their suboptimality when gradient descent is used. To address the suboptimality issue, we propose the Transformation And Distillation (TAD) framework, which reformulates a multi-agent MDP as a special single-agent MDP with a sequential structure and enables decentralized execution by distilling the learned policy on the derived "single-agent" MDP. The approach is a two-stage learning paradigm that addresses the optimization problem in cooperative MARL, providing optimality guarantee with decent execution performance. Empirically, we implement TAD-PPO based on PPO, which can theoretically perform optimal policy learning in the finite multi-agent MDPs and shows significant outperformance on a large set of cooperative multi-agent tasks, from matrix game, hallway task, to StarCraft II, and football game.

Zheng Zhang, Yingsheng Ji, Jiachen Shen et al.

Risk assessment is a substantial problem for financial institutions that has been extensively studied both for its methodological richness and its various practical applications. With the expansion of inclusive finance, recent attentions are paid to micro and small-sized enterprises (MSEs). Compared with large companies, MSEs present a higher exposure rate to default owing to their insecure financial stability. Conventional efforts learn classifiers from historical data with elaborate feature engineering. However, the main obstacle for MSEs involves severe deficiency in credit-related information, which may degrade the performance of prediction. Besides, financial activities have diverse explicit and implicit relations, which have not been fully exploited for risk judgement in commercial banks. In particular, the observations on real data show that various relationships between company users have additional power in financial risk analysis. In this paper, we consider a graph of banking data, and propose a novel HIDAM model for the purpose. Specifically, we attempt to incorporate heterogeneous information network with rich attributes on multi-typed nodes and links for modeling the scenario of business banking service. To enhance feature representation of MSEs, we extract interactive information through meta-paths and fully exploit path information. Furthermore, we devise a hierarchical attention mechanism respectively to learn the importance of contents inside each meta-path and the importance of different metapahs. Experimental results verify that HIDAM outperforms state-of-the-art competitors on real-world banking data.

Haoyu Ma, Yushu Chen, Wenlai Zhao et al. · tsinghua

Time series foundation models have demonstrated strong performance in zero-shot learning, making them well-suited for predicting rapidly evolving patterns in real-world applications where relevant training data are scarce. However, most of these models rely on the Transformer architecture, which incurs quadratic complexity as input length increases. To address this, we introduce TSMamba, a linear-complexity foundation model for time series forecasting built on the Mamba architecture. The model captures temporal dependencies through both forward and backward Mamba encoders, achieving high prediction accuracy. To reduce reliance on large datasets and lower training costs, TSMamba employs a two-stage transfer learning process that leverages pretrained Mamba LLMs, allowing effective time series modeling with a moderate training set. In the first stage, the forward and backward backbones are optimized via patch-wise autoregressive prediction; in the second stage, the model trains a prediction head and refines other components for long-term forecasting. While the backbone assumes channel independence to manage varying channel numbers across datasets, a channel-wise compressed attention module is introduced to capture cross-channel dependencies during fine-tuning on specific multivariate datasets. Experiments show that TSMamba's zero-shot performance is comparable to state-of-the-art time series foundation models, despite using significantly less training data. It also achieves competitive or superior full-shot performance compared to task-specific prediction models. The code will be made publicly available.

Yushu Chen, Shengzhuo Liu, Jinzhe Yang et al.

In order to enhance the performance of Transformer models for long-term multivariate forecasting while minimizing computational demands, this paper introduces the Joint Time-Frequency Domain Transformer (JTFT). JTFT combines time and frequency domain representations to make predictions. The frequency domain representation efficiently extracts multi-scale dependencies while maintaining sparsity by utilizing a small number of learnable frequencies. Simultaneously, the time domain (TD) representation is derived from a fixed number of the most recent data points, strengthening the modeling of local relationships and mitigating the effects of non-stationarity. Importantly, the length of the representation remains independent of the input sequence length, enabling JTFT to achieve linear computational complexity. Furthermore, a low-rank attention layer is proposed to efficiently capture cross-dimensional dependencies, thus preventing performance degradation resulting from the entanglement of temporal and channel-wise modeling. Experimental results on six real-world datasets demonstrate that JTFT outperforms state-of-the-art baselines in predictive performance.

Zichuan Lin, Garrett Thomas, Guangwen Yang et al.

Meta-reinforcement learning (meta-RL) aims to learn from multiple training tasks the ability to adapt efficiently to unseen test tasks. Despite the success, existing meta-RL algorithms are known to be sensitive to the task distribution shift. When the test task distribution is different from the training task distribution, the performance may degrade significantly. To address this issue, this paper proposes Model-based Adversarial Meta-Reinforcement Learning (AdMRL), where we aim to minimize the worst-case sub-optimality gap -- the difference between the optimal return and the return that the algorithm achieves after adaptation -- across all tasks in a family of tasks, with a model-based approach. We propose a minimax objective and optimize it by alternating between learning the dynamics model on a fixed task and finding the adversarial task for the current model -- the task for which the policy induced by the model is maximally suboptimal. Assuming the family of tasks is parameterized, we derive a formula for the gradient of the suboptimality with respect to the task parameters via the implicit function theorem, and show how the gradient estimator can be efficiently implemented by the conjugate gradient method and a novel use of the REINFORCE estimator. We evaluate our approach on several continuous control benchmarks and demonstrate its efficacy in the worst-case performance over all tasks, the generalization power to out-of-distribution tasks, and in training and test time sample efficiency, over existing state-of-the-art meta-RL algorithms.

Mingzhen Li, Yi Liu, Xiaoyan Liu et al.

The difficulty of deploying various deep learning (DL) models on diverse DL hardware has boosted the research and development of DL compilers in the community. Several DL compilers have been proposed from both industry and academia such as Tensorflow XLA and TVM. Similarly, the DL compilers take the DL models described in different DL frameworks as input, and then generate optimized codes for diverse DL hardware as output. However, none of the existing survey has analyzed the unique design architecture of the DL compilers comprehensively. In this paper, we perform a comprehensive survey of existing DL compilers by dissecting the commonly adopted design in details, with emphasis on the DL oriented multi-level IRs, and frontend/backend optimizations. Specifically, we provide a comprehensive comparison among existing DL compilers from various aspects. In addition, we present detailed analysis on the design of multi-level IRs and illustrate the commonly adopted optimization techniques. Finally, several insights are highlighted as the potential research directions of DL compiler. This is the first survey paper focusing on the design architecture of DL compilers, which we hope can pave the road for future research towards DL compiler.

Zichuan Lin, Li Zhao, Derek Yang et al.

Many reinforcement learning (RL) tasks have specific properties that can be leveraged to modify existing RL algorithms to adapt to those tasks and further improve performance, and a general class of such properties is the multiple reward channel. In those environments the full reward can be decomposed into sub-rewards obtained from different channels. Existing work on reward decomposition either requires prior knowledge of the environment to decompose the full reward, or decomposes reward without prior knowledge but with degraded performance. In this paper, we propose Distributional Reward Decomposition for Reinforcement Learning (DRDRL), a novel reward decomposition algorithm which captures the multiple reward channel structure under distributional setting. Empirically, our method captures the multi-channel structure and discovers meaningful reward decomposition, without any requirements on prior knowledge. Consequently, our agent achieves better performance than existing methods on environments with multiple reward channels.

Yushu Chen, Hao Jing, Wenlai Zhao et al.

We present the remote stochastic gradient (RSG) method, which computes the gradients at configurable remote observation points, in order to improve the convergence rate and suppress gradient noise at the same time for different curvatures. RSG is further combined with adaptive methods to construct ARSG for acceleration. The method is efficient in computation and memory, and is straightforward to implement. We analyze the convergence properties by modeling the training process as a dynamic system, which provides a guideline to select the configurable observation factor without grid search. ARSG yields $O(1/\sqrt{T})$ convergence rate in non-convex settings, that can be further improved to $O(\log(T)/T)$ in strongly convex settings. Numerical experiments demonstrate that ARSG achieves both faster convergence and better generalization, compared with popular adaptive methods, such as ADAM, NADAM, AMSGRAD, and RANGER for the tested problems. In particular, for training ResNet-50 on ImageNet, ARSG outperforms ADAM in convergence speed and meanwhile it surpasses SGD in generalization.

Mingzhen Li, Changxi Liu, Jianjin Liao et al.

The flourish of deep learning frameworks and hardware platforms has been demanding an efficient compiler that can shield the diversity in both software and hardware in order to provide application portability. Among the existing deep learning compilers, TVM is well known for its efficiency in code generation and optimization across diverse hardware devices. In the meanwhile, the Sunway many-core processor renders itself as a competitive candidate for its attractive computational power in both scientific computing and deep learning workloads. This paper combines the trends in these two directions. Specifically, we propose swTVM that extends the original TVM to support ahead-of-time compilation for architecture requiring cross-compilation such as Sunway. In addition, we leverage the architecture features during the compilation such as core group for massive parallelism, DMA for high bandwidth memory transfer and local device memory for data locality, in order to generate efficient codes for deep learning workloads on Sunway. The experiment results show that the codes generated by swTVM achieves 1.79x on average compared to the state-of-the-art deep learning framework on Sunway, across six representative benchmarks. This work is the first attempt from the compiler perspective to bridge the gap of deep learning and Sunway processor particularly with productivity and efficiency in mind. We believe this work will encourage more people to embrace the power of deep learning and Sunway many-core processor.

Jiarui Fang, Liandeng Li, Haohuan Fu et al.

This paper reports our efforts on swCaffe, a highly efficient parallel framework for accelerating deep neural networks (DNNs) training on Sunway TaihuLight, the current fastest supercomputer in the world that adopts a unique many-core heterogeneous architecture, with 40,960 SW26010 processors connected through a customized communication network. First, we point out some insightful principles to fully exploit the performance of the innovative many-core architecture. Second, we propose a set of optimization strategies for redesigning a variety of neural network layers based on Caffe. Third, we put forward a topology-aware parameter synchronization scheme to scale the synchronous Stochastic Gradient Descent (SGD) method to multiple processors efficiently. We evaluate our framework by training a variety of widely used neural networks with the ImageNet dataset. On a single node, swCaffe can achieve 23\%\~{}119\% overall performance compared with Caffe running on K40m GPU. As compared with the Caffe on CPU, swCaffe runs 3.04\~{}7.84x faster on all the networks. Finally, we present the scalability of swCaffe for the training of ResNet-50 and AlexNet on the scale of 1024 nodes.

Jiarui Fang, Haohuan Fu, Guangwen Yang et al.

Data parallelism has become a dominant method to scale Deep Neural Network (DNN) training across multiple nodes. Since synchronizing a large number of gradients of the local model can be a bottleneck for large-scale distributed training, compressing communication data has gained widespread attention recently. Among several recent proposed compression algorithms, Residual Gradient Compression (RGC) is one of the most successful approaches---it can significantly compress the transmitting message size (0.1\% of the gradient size) of each node and still achieve correct accuracy and the same convergence speed. However, the literature on compressing deep networks focuses almost exclusively on achieving good theoretical compression rate, while the efficiency of RGC in real distributed implementation has been less investigated. In this paper, we develop an RGC-based system that is able to reduce the end-to-end training time on real-world multi-GPU systems. Our proposed design called RedSync, which introduces a set of optimizations to reduce communication bandwidth requirement while introducing limited overhead. We evaluate the performance of RedSync on two different multiple GPU platforms, including 128 GPUs of a supercomputer and an 8-GPU server. Our test cases include image classification tasks on Cifar10 and ImageNet, and language modeling tasks on Penn Treebank and Wiki2 datasets. For DNNs featured with high communication to computation ratio, which have long been considered with poor scalability, RedSync brings significant performance improvements.

Zichuan Lin, Tianqi Zhao, Guangwen Yang et al.

Reinforcement learning (RL) algorithms have made huge progress in recent years by leveraging the power of deep neural networks (DNN). Despite the success, deep RL algorithms are known to be sample inefficient, often requiring many rounds of interaction with the environments to obtain satisfactory performance. Recently, episodic memory based RL has attracted attention due to its ability to latch on good actions quickly. In this paper, we present a simple yet effective biologically inspired RL algorithm called Episodic Memory Deep Q-Networks (EMDQN), which leverages episodic memory to supervise an agent during training. Experiments show that our proposed method can lead to better sample efficiency and is more likely to find good policies. It only requires 1/5 of the interactions of DQN to achieve many state-of-the-art performances on Atari games, significantly outperforming regular DQN and other episodic memory based RL algorithms.

Hao Zhang, Shizhen Xu, Graham Neubig et al.

Recent deep learning (DL) models have moved beyond static network architectures to dynamic ones, handling data where the network structure changes every example, such as sequences of variable lengths, trees, and graphs. Existing dataflow-based programming models for DL---both static and dynamic declaration---either cannot readily express these dynamic models, or are inefficient due to repeated dataflow graph construction and processing, and difficulties in batched execution. We present Cavs, a vertex-centric programming interface and optimized system implementation for dynamic DL models. Cavs represents dynamic network structure as a static vertex function $\mathcal{F}$ and a dynamic instance-specific graph $\mathcal{G}$, and performs backpropagation by scheduling the execution of $\mathcal{F}$ following the dependencies in $\mathcal{G}$. Cavs bypasses expensive graph construction and preprocessing overhead, allows for the use of static graph optimization techniques on pre-defined operations in $\mathcal{F}$, and naturally exposes batched execution opportunities over different graphs. Experiments comparing Cavs to two state-of-the-art frameworks for dynamic NNs (TensorFlow Fold and DyNet) demonstrate the efficacy of this approach: Cavs achieves a near one order of magnitude speedup on training of various dynamic NN architectures, and ablations demonstrate the contribution of our proposed batching and memory management strategies.