We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1. DeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT) as a preliminary step, demonstrates remarkable reasoning capabilities. Through RL, DeepSeek-R1-Zero naturally emerges with numerous powerful and intriguing reasoning behaviors. However, it encounters challenges such as poor readability, and language mixing. To address these issues and further enhance reasoning performance, we introduce DeepSeek-R1, which incorporates multi-stage training and cold-start data before RL. DeepSeek-R1 achieves performance comparable to OpenAI-o1-1217 on reasoning tasks. To support the research community, we open-source DeepSeek-R1-Zero, DeepSeek-R1, and six dense models (1.5B, 7B, 8B, 14B, 32B, 70B) distilled from DeepSeek-R1 based on Qwen and Llama.

We present DeepSeek-V3, a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token. To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly validated in DeepSeek-V2. Furthermore, DeepSeek-V3 pioneers an auxiliary-loss-free strategy for load balancing and sets a multi-token prediction training objective for stronger performance. We pre-train DeepSeek-V3 on 14.8 trillion diverse and high-quality tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages to fully harness its capabilities. Comprehensive evaluations reveal that DeepSeek-V3 outperforms other open-source models and achieves performance comparable to leading closed-source models. Despite its excellent performance, DeepSeek-V3 requires only 2.788M H800 GPU hours for its full training. In addition, its training process is remarkably stable. Throughout the entire training process, we did not experience any irrecoverable loss spikes or perform any rollbacks.

Artificial intelligence generated content (AIGC), a rapidly advancing technology, is transforming content creation across domains, such as text, images, audio, and video. Its growing potential has attracted more and more researchers and investors to explore and expand its possibilities. This review traces AIGC's evolution through four developmental milestones-ranging from early rule-based systems to modern transfer learning models-within a unified framework that highlights how each milestone contributes uniquely to content generation. In particular, the paper employs a common example across all milestones to illustrate the capabilities and limitations of methods within each phase, providing a consistent evaluation of AIGC methodologies and their development. Furthermore, this paper addresses critical challenges associated with AIGC and proposes actionable strategies to mitigate them. This study aims to guide researchers and practitioners in selecting and optimizing AIGC models to enhance the quality and efficiency of content creation across diverse domains.

In the search for more effective education, there is a widespread effort to develop better approaches to personalize student education. Unassisted, educators often do not have time or resources to personally support every student in a given classroom. Motivated by this issue, and by recent advancements in artificial intelligence, this paper presents a general-purpose framework for personalized student support, applicable to any virtual educational system such as a serious game or an intelligent tutoring system. To fit any educational situation, we apply ontologies for their semantic organization, combining them with data collection considerations and multi-agent reinforcement learning. The result is a modular system that can be adapted to any virtual educational software to provide useful personalized assistance to students.

Due to the effective performance of multi-scale feature fusion, Path Aggregation FPN (PAFPN) is widely employed in YOLO detectors. However, it cannot efficiently and adaptively integrate high-level semantic information with low-level spatial information simultaneously. We propose a new model named MAF-YOLO in this paper, which is a novel object detection framework with a versatile neck named Multi-Branch Auxiliary FPN (MAFPN). Within MAFPN, the Superficial Assisted Fusion (SAF) module is designed to combine the output of the backbone with the neck, preserving an optimal level of shallow information to facilitate subsequent learning. Meanwhile, the Advanced Assisted Fusion (AAF) module deeply embedded within the neck conveys a more diverse range of gradient information to the output layer. Furthermore, our proposed Re-parameterized Heterogeneous Efficient Layer Aggregation Network (RepHELAN) module ensures that both the overall model architecture and convolutional design embrace the utilization of heterogeneous large convolution kernels. Therefore, this guarantees the preservation of information related to small targets while simultaneously achieving the multi-scale receptive field. Finally, taking the nano version of MAF-YOLO for example, it can achieve 42.4% AP on COCO with only 3.76M learnable parameters and 10.51G FLOPs, and approximately outperforms YOLOv8n by about 5.1%. The source code of this work is available at: https://github.com/yang-0201/MAF-YOLO.

We present DeepSeek-V2, a strong Mixture-of-Experts (MoE) language model characterized by economical training and efficient inference. It comprises 236B total parameters, of which 21B are activated for each token, and supports a context length of 128K tokens. DeepSeek-V2 adopts innovative architectures including Multi-head Latent Attention (MLA) and DeepSeekMoE. MLA guarantees efficient inference through significantly compressing the Key-Value (KV) cache into a latent vector, while DeepSeekMoE enables training strong models at an economical cost through sparse computation. Compared with DeepSeek 67B, DeepSeek-V2 achieves significantly stronger performance, and meanwhile saves 42.5% of training costs, reduces the KV cache by 93.3%, and boosts the maximum generation throughput to 5.76 times. We pretrain DeepSeek-V2 on a high-quality and multi-source corpus consisting of 8.1T tokens, and further perform Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) to fully unlock its potential. Evaluation results show that, even with only 21B activated parameters, DeepSeek-V2 and its chat versions still achieve top-tier performance among open-source models.

Noise is usually regarded as adversarial to extract the effective dynamics from time series, such that the conventional data-driven approaches usually aim at learning the dynamics by mitigating the noisy effect. However, noise can have a functional role of driving transitions between stable states underlying many natural and engineered stochastic dynamics. To capture such stochastic transitions from data, we find that leveraging a machine learning model, reservoir computing as a type of recurrent neural network, can learn noise-induced transitions. We develop a concise training protocol for tuning hyperparameters, with a focus on a pivotal hyperparameter controlling the time scale of the reservoir dynamics. The trained model generates accurate statistics of transition time and the number of transitions. The approach is applicable to a wide class of systems, including a bistable system under a double-well potential, with either white noise or colored noise. It is also aware of the asymmetry of the double-well potential, the rotational dynamics caused by non-detailed balance, and transitions in multi-stable systems. For the experimental data of protein folding, it learns the transition time between folded states, providing a possibility of predicting transition statistics from a small dataset. The results demonstrate the capability of machine-learning methods in capturing noise-induced phenomena.

Multi-agent reinforcement learning (MARL) is a widely used Artificial Intelligence (AI) technique. However, current studies and applications need to address its scalability, non-stationarity, and trustworthiness. This paper aims to review methods and applications and point out research trends and visionary prospects for the next decade. First, this paper summarizes the basic methods and application scenarios of MARL. Second, this paper outlines the corresponding research methods and their limitations on safety, robustness, generalization, and ethical constraints that need to be addressed in the practical applications of MARL. In particular, we believe that trustworthy MARL will become a hot research topic in the next decade. In addition, we suggest that considering human interaction is essential for the practical application of MARL in various societies. Therefore, this paper also analyzes the challenges while MARL is applied to human-machine interaction.

International Centre for Theoretical Physics Asia-Pacific, Beijing/Hangzhou, China The stochastic reaction network in which chemical species evolve through a set of reactions is widely used to model stochastic processes in physics, chemistry and biology. To characterize the evolving joint probability distribution in the state space of species counts requires solving a system of ordinary differential equations, the chemical master equation, where the size of the counting state space increases exponentially with the type of species, making it challenging to investigate the stochastic reaction network. Here, we propose a machine-learning approach using the variational autoregressive network to solve the chemical master equation. Training the autoregressive network employs the policy gradient algorithm in the reinforcement learning framework, which does not require any data simulated in prior by another method. Different from simulating single trajectories, the approach tracks the time evolution of the joint probability distribution, and supports direct sampling of configurations and computing their normalized joint probabilities. We apply the approach to representative examples in physics and biology, and demonstrate that it accurately generates the probability distribution over time. The variational autoregressive network exhibits a plasticity in representing the multimodal distribution, cooperates with the conservation law, enables time-dependent reaction rates, and is efficient for high-dimensional reaction networks with allowing a flexible upper count limit. The results suggest a general approach to investigate stochastic reaction networks based on modern machine learning. The stochastic reaction network is a standard model for stochastic processes in physics [1], chemistry [2, 3], biology [4], and ecology [5, 6]. Representative examples include the birth-death process [7], the model of spontaneous asymmetric synthesis [8], and gene regulatory networks [9].