Naifu Zhang

Naifu Zhang, Meixia Tao, Jia Wang et al.

One of the main focuses in distributed learning is communication efficiency, since model aggregation at each round of training can consist of millions to billions of parameters. Several model compression methods, such as gradient quantization and sparsification, have been proposed to improve the communication efficiency of model aggregation. However, the information-theoretic minimum communication cost for a given distortion of gradient estimators is still unknown. In this paper, we study the fundamental limit of communication cost of model aggregation in distributed learning from a rate-distortion perspective. By formulating the model aggregation as a vector Gaussian CEO problem, we derive the rate region bound and sum-rate-distortion function for the model aggregation problem, which reveals the minimum communication rate at a particular gradient distortion upper bound. We also analyze the communication cost at each iteration and total communication cost based on the sum-rate-distortion function with the gradient statistics of real-world datasets. It is found that the communication gain by exploiting the correlation between worker nodes is significant for SignSGD, and a high distortion of gradient estimator can achieve low total communication cost in gradient compression.

Shaojie Shi, Zhengyu Shi, Lingran Zheng et al.

Causal inference in social science relies on end-to-end, intervention-centered research-design reasoning grounded in real-world policy interventions, but current benchmarks fail to evaluate this capability of large language models (LLMs). We present InterveneBench, a benchmark designed to assess such reasoning in realistic social settings. Each instance in InterveneBench is derived from an empirical social science study and requires models to reason about policy interventions and identification assumptions without access to predefined causal graphs or structural equations. InterveneBench comprises 744 peer-reviewed studies across diverse policy domains. Experimental results show that state-of-the-art LLMs struggle under this setting. To address this limitation, we further propose a multi-agent framework, STRIDES. It achieves significant performance improvements over state-of-the-art reasoning models. Our code and data are available at https://github.com/Sii-yuning/STRIDES.

Naifu Zhang, Wei Tao, Xi Xiao et al.

In recent years, Vision-Language-Action (VLA) models in embodied intelligence have developed rapidly. However, existing adversarial attack methods require costly end-to-end training and often generate noticeable perturbation patches. To address these limitations, we propose ADVLA, a framework that directly applies adversarial perturbations on features projected from the visual encoder into the textual feature space. ADVLA efficiently disrupts downstream action predictions under low-amplitude constraints, and attention guidance allows the perturbations to be both focused and sparse. We introduce three strategies that enhance sensitivity, enforce sparsity, and concentrate perturbations. Experiments demonstrate that under an $L_{\infty}=4/255$ constraint, ADVLA combined with Top-K masking modifies less than 10% of the patches while achieving an attack success rate of nearly 100%. The perturbations are concentrated on critical regions, remain almost imperceptible in the overall image, and a single-step iteration takes only about 0.06 seconds, significantly outperforming conventional patch-based attacks. In summary, ADVLA effectively weakens downstream action predictions of VLA models under low-amplitude and locally sparse conditions, avoiding the high training costs and conspicuous perturbations of traditional patch attacks, and demonstrates unique effectiveness and practical value for attacking VLA feature spaces.

Yichun Yin, Wenyong Huang, Kaikai Song et al.

We present Pangu Ultra, a Large Language Model (LLM) with 135 billion parameters and dense Transformer modules trained on Ascend Neural Processing Units (NPUs). Although the field of LLM has been witnessing unprecedented advances in pushing the scale and capability of LLM in recent years, training such a large-scale model still involves significant optimization and system challenges. To stabilize the training process, we propose depth-scaled sandwich normalization, which effectively eliminates loss spikes during the training process of deep models. We pre-train our model on 13.2 trillion diverse and high-quality tokens and further enhance its reasoning capabilities during post-training. To perform such large-scale training efficiently, we utilize 8,192 Ascend NPUs with a series of system optimizations. Evaluations on multiple diverse benchmarks indicate that Pangu Ultra significantly advances the state-of-the-art capabilities of dense LLMs such as Llama 405B and Mistral Large 2, and even achieves competitive results with DeepSeek-R1, whose sparse model structure contains much more parameters. Our exploration demonstrates that Ascend NPUs are capable of efficiently and effectively training dense models with more than 100 billion parameters. Our model and system will be available for our commercial customers.

Yehui Tang, Yichun Yin, Yaoyuan Wang et al.

Sparse large language models (LLMs) with Mixture of Experts (MoE) and close to a trillion parameters are dominating the realm of most capable language models. However, the massive model scale poses significant challenges for the underlying software and hardware systems. In this paper, we aim to uncover a recipe to harness such scale on Ascend NPUs. The key goals are better usage of the computing resources under the dynamic sparse model structures and materializing the expected performance gain on the actual hardware. To select model configurations suitable for Ascend NPUs without repeatedly running the expensive experiments, we leverage simulation to compare the trade-off of various model hyperparameters. This study led to Pangu Ultra MoE, a sparse LLM with 718 billion parameters, and we conducted experiments on the model to verify the simulation results. On the system side, we dig into Expert Parallelism to optimize the communication between NPU devices to reduce the synchronization overhead. We also optimize the memory efficiency within the devices to further reduce the parameter and activation management overhead. In the end, we achieve an MFU of 30.0% when training Pangu Ultra MoE, with performance comparable to that of DeepSeek R1, on 6K Ascend NPUs, and demonstrate that the Ascend system is capable of harnessing all the training stages of the state-of-the-art language models. Extensive experiments indicate that our recipe can lead to efficient training of large-scale sparse language models with MoE. We also study the behaviors of such models for future reference.

Naifu Zhang, Meixia Tao, Jia Wang

One of the main focus in federated learning (FL) is the communication efficiency since a large number of participating edge devices send their updates to the edge server at each round of the model training. Existing works reconstruct each model update from edge devices and implicitly assume that the local model updates are independent over edge devices. In FL, however, the model update is an indirect multi-terminal source coding problem, also called as the CEO problem where each edge device cannot observe directly the gradient that is to be reconstructed at the decoder, but is rather provided only with a noisy version. The existing works do not leverage the redundancy in the information transmitted by different edges. This paper studies the rate region for the indirect multiterminal source coding problem in FL. The goal is to obtain the minimum achievable rate at a particular upper bound of gradient variance. We obtain the rate region for the quadratic vector Gaussian CEO problem under unbiased estimator and derive an explicit formula of the sum-rate-distortion function in the special case where gradient are identical over edge device and dimension. Finally, we analyse communication efficiency of convex Minibatched SGD and non-convex Minibatched SGD based on the sum-rate-distortion function, respectively.

Nicholas Capel, Naifu Zhang

There have been attempts in reinforcement learning to exploit a priori knowledge about the structure of the system. This paper proposes a hybrid reinforcement learning controller which dynamically interpolates a model-based linear controller and an arbitrary differentiable policy. The linear controller is designed based on local linearised model knowledge, and stabilises the system in a neighbourhood about an operating point. The coefficients of interpolation between the two controllers are determined by a scaled distance function measuring the distance between the current state and the operating point. The overall hybrid controller is proven to maintain the stability guarantee around the neighborhood of the operating point and still possess the universal function approximation property of the arbitrary non-linear policy. Learning has been done on both model-based (PILCO) and model-free (DDPG) frameworks. Simulation experiments performed in OpenAI gym demonstrate stability and robustness of the proposed hybrid controller. This paper thus introduces a principled method allowing for the direct importing of control methodology into reinforcement learning.

Yuhong Deng, Di Guo, Xiaofeng Guo et al.

In this paper, we propose a novel task, Manipulation Question Answering (MQA), where the robot performs manipulation actions to change the environment in order to answer a given question. To solve this problem, a framework consisting of a QA module and a manipulation module is proposed. For the QA module, we adopt the method for the Visual Question Answering (VQA) task. For the manipulation module, a Deep Q Network (DQN) model is designed to generate manipulation actions for the robot to interact with the environment. We consider the situation where the robot continuously manipulating objects inside a bin until the answer to the question is found. Besides, a novel dataset that contains a variety of object models, scenarios and corresponding question-answer pairs is established in a simulation environment. Extensive experiments have been conducted to validate the effectiveness of the proposed framework.

Naifu Zhang, Meixia Tao

Federated learning (FL) is a promising technique that enables many edge devices to train a machine learning model collaboratively in wireless networks. By exploiting the superposition nature of wireless waveforms, over-the-air computation (AirComp) can accelerate model aggregation and hence facilitate communication-efficient FL. Due to channel fading, power control is crucial in AirComp. Prior works assume that the signals to be aggregated from each device, i.e., local gradients have identical statistics. In FL, however, gradient statistics vary over both training iterations and feature dimensions, and are unknown in advance. This paper studies the power control problem for over-the-air FL by taking gradient statistics into account. The goal is to minimize the aggregation error by optimizing the transmit power at each device subject to peak power constraints. We obtain the optimal policy in closed form when gradient statistics are given. Notably, we show that the optimal transmit power is continuous and monotonically decreases with the squared multivariate coefficient of variation (SMCV) of gradient vectors. We then propose a method to estimate gradient statistics with negligible communication cost. Experimental results demonstrate that the proposed gradient-statistics-aware power control achieves higher test accuracy than the existing schemes for a wide range of scenarios.