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.

Growing scale of recommender systems require extensive tuning to respond to market dynamics and system changes. We address the challenge of tuning a large-scale ads recommendation platform with multiple continuous parameters influencing key performance indicators (KPIs). Traditional methods like open-box Monte Carlo simulators, while accurate, are computationally expensive due to the high cost of evaluating numerous parameter settings. To mitigate this, we propose a hybrid approach Simulator-Guided Importance Sampling (SGIS) that combines open-box simulation with importance sampling (IS). SGIS leverages the strengths of both techniques: it performs a coarse enumeration over the parameter space to identify promising initial settings and then uses IS to iteratively refine these settings. This approach significantly reduces computational costs while maintaining high accuracy in KPI estimation. We demonstrate the effectiveness of SGIS through simulations as well as real-world experiments, showing that it achieves substantial improvements in KPIs with lower computational overhead compared to traditional methods.

This paper introduces Adaptive Mixture Importance Sampling (AMIS) as a novel approach for optimizing key performance indicators (KPIs) in large-scale recommender systems, such as online ad auctions. Traditional importance sampling (IS) methods face challenges in dynamic environments, particularly in navigating through complexities of multi-modal landscapes and avoiding entrapment in local optima for the optimization task. Instead of updating importance weights and mixing samples across iterations, as in canonical adaptive IS and multiple IS, our AMIS framework leverages a mixture distribution as the proposal distribution and dynamically adjusts both the mixture parameters and their mixing rates at each iteration, thereby enhancing search diversity and efficiency. Through extensive offline simulations, we demonstrate that AMIS significantly outperforms simple Gaussian Importance Sampling (GIS), particularly in noisy environments. Moreover, our approach is validated in real-world scenarios through online A/B experiments on a major search engine, where AMIS consistently identifies optimal tuning points that are more likely to be adopted as mainstream configurations. These findings indicate that AMIS enhances convergence in noisy environments, leading to more accurate and reliable decision-making in the context of importance sampling off-policy estimators.

Personalized Federated Learning (PFL) is proposed to find the greatest personalized models for each client. To avoid the central failure and communication bottleneck in the server-based FL, we concentrate on the Decentralized Personalized Federated Learning (DPFL) that performs distributed model training in a Peer-to-Peer (P2P) manner. Most personalized works in DPFL are based on undirected and symmetric topologies, however, the data, computation and communication resources heterogeneity result in large variances in the personalized models, which lead the undirected aggregation to suboptimal personalized performance and unguaranteed convergence. To address these issues, we propose a directed collaboration DPFL framework by incorporating stochastic gradient push and partial model personalized, called \textbf{D}ecentralized \textbf{Fed}erated \textbf{P}artial \textbf{G}radient \textbf{P}ush (\textbf{DFedPGP}). It personalizes the linear classifier in the modern deep model to customize the local solution and learns a consensus representation in a fully decentralized manner. Clients only share gradients with a subset of neighbors based on the directed and asymmetric topologies, which guarantees flexible choices for resource efficiency and better convergence. Theoretically, we show that the proposed DFedPGP achieves a superior convergence rate of $\mathcal{O}(\frac{1}{\sqrt{T}})$ in the general non-convex setting, and prove the tighter connectivity among clients will speed up the convergence. The proposed method achieves state-of-the-art (SOTA) accuracy in both data and computation heterogeneity scenarios, demonstrating the efficiency of the directed collaboration and partial gradient push.

With the blowout development of pre-trained models (PTMs), the efficient tuning of these models for diverse downstream applications has emerged as a pivotal research concern. Although recent investigations into prompt tuning have provided promising avenues, three salient challenges persist: (1) memory constraint: the continuous growth in the size of open-source PTMs renders fine-tuning, even a fraction of their parameters, challenging for many practitioners. (2) model privacy: existing PTMs often function as public API services, with their parameters inaccessible for effective or tailored fine-tuning. (3) data privacy: the fine-tuning of PTMs necessitates high-quality datasets, which are typically localized and not shared to public. To optimally harness each local dataset while navigating memory constraints and preserving privacy, we propose Federated Black-Box Prompt Tuning (Fed-BBPT). This innovative approach eschews reliance on parameter architectures and private dataset access, instead capitalizing on a central server that aids local users in collaboratively training a prompt generator through regular aggregation. Local users leverage API-driven learning via a zero-order optimizer, obviating the need for PTM deployment. Relative to extensive fine-tuning, Fed-BBPT proficiently sidesteps memory challenges tied to PTM storage and fine-tuning on local machines, tapping into comprehensive, high-quality, yet private training datasets. A thorough evaluation across 40 datasets spanning CV and NLP tasks underscores the robustness of our proposed model.

To defend the inference attacks and mitigate the sensitive information leakages in Federated Learning (FL), client-level Differentially Private FL (DPFL) is the de-facto standard for privacy protection by clipping local updates and adding random noise. However, existing DPFL methods tend to make a sharper loss landscape and have poorer weight perturbation robustness, resulting in severe performance degradation. To alleviate these issues, we propose a novel DPFL algorithm named DP-FedSAM, which leverages gradient perturbation to mitigate the negative impact of DP. Specifically, DP-FedSAM integrates Sharpness Aware Minimization (SAM) optimizer to generate local flatness models with better stability and weight perturbation robustness, which results in the small norm of local updates and robustness to DP noise, thereby improving the performance. From the theoretical perspective, we analyze in detail how DP-FedSAM mitigates the performance degradation induced by DP. Meanwhile, we give rigorous privacy guarantees with R\'enyi DP and present the sensitivity analysis of local updates. At last, we empirically confirm that our algorithm achieves state-of-the-art (SOTA) performance compared with existing SOTA baselines in DPFL. Code is available at https://github.com/YMJS-Irfan/DP-FedSAM

To mitigate the privacy leakages and communication burdens of Federated Learning (FL), decentralized FL (DFL) discards the central server and each client only communicates with its neighbors in a decentralized communication network. However, existing DFL suffers from high inconsistency among local clients, which results in severe distribution shift and inferior performance compared with centralized FL (CFL), especially on heterogeneous data or sparse communication topology. To alleviate this issue, we propose two DFL algorithms named DFedSAM and DFedSAM-MGS to improve the performance of DFL. Specifically, DFedSAM leverages gradient perturbation to generate local flat models via Sharpness Aware Minimization (SAM), which searches for models with uniformly low loss values. DFedSAM-MGS further boosts DFedSAM by adopting Multiple Gossip Steps (MGS) for better model consistency, which accelerates the aggregation of local flat models and better balances communication complexity and generalization. Theoretically, we present improved convergence rates $\small \mathcal{O}\big(\frac{1}{\sqrt{KT}}+\frac{1}{T}+\frac{1}{K^{1/2}T^{3/2}(1-\lambda)^2}\big)$ and $\small \mathcal{O}\big(\frac{1}{\sqrt{KT}}+\frac{1}{T}+\frac{\lambda^Q+1}{K^{1/2}T^{3/2}(1-\lambda^Q)^2}\big)$ in non-convex setting for DFedSAM and DFedSAM-MGS, respectively, where $1-\lambda$ is the spectral gap of gossip matrix and $Q$ is the number of MGS. Empirically, our methods can achieve competitive performance compared with CFL methods and outperform existing DFL methods.

Personalized federated learning (PFL) aims to produce the greatest personalized model for each client to face an insurmountable problem - data heterogeneity in real FL systems. However, almost all existing works have to face large communication burdens and the risk of disruption if the central server fails. Only limited efforts have been used in a decentralized way but still suffers from inferior representation ability due to sharing the full model with its neighbors. Therefore, in this paper, we propose a personalized FL framework with a decentralized partial model training called DFedAlt. It personalizes the "right" components in the modern deep models by alternately updating the shared and personal parameters to train partially personalized models in a peer-to-peer manner. To further promote the shared parameters aggregation process, we propose DFedSalt integrating the local Sharpness Aware Minimization (SAM) optimizer to update the shared parameters. It adds proper perturbation in the direction of the gradient to overcome the shared model inconsistency across clients. Theoretically, we provide convergence analysis of both algorithms in the general non-convex setting for decentralized partial model training in PFL. Our experiments on several real-world data with various data partition settings demonstrate that (i) decentralized training is more suitable for partial personalization, which results in state-of-the-art (SOTA) accuracy compared with the SOTA PFL baselines; (ii) the shared parameters with proper perturbation make partial personalized FL more suitable for decentralized training, where DFedSalt achieves most competitive performance.

To defend the inference attacks and mitigate the sensitive information leakages in Federated Learning (FL), client-level Differentially Private FL (DPFL) is the de-facto standard for privacy protection by clipping local updates and adding random noise. However, existing DPFL methods tend to make a sharp loss landscape and have poor weight perturbation robustness, resulting in severe performance degradation. To alleviate these issues, we propose a novel DPFL algorithm named DP-FedSAM, which leverages gradient perturbation to mitigate the negative impact of DP. Specifically, DP-FedSAM integrates Sharpness Aware Minimization (SAM) optimizer to generate local flatness models with improved stability and weight perturbation robustness, which results in the small norm of local updates and robustness to DP noise, thereby improving the performance. To further reduce the magnitude of random noise while achieving better performance, we propose DP-FedSAM-$top_k$ by adopting the local update sparsification technique. From the theoretical perspective, we present the convergence analysis to investigate how our algorithms mitigate the performance degradation induced by DP. Meanwhile, we give rigorous privacy guarantees with R\'enyi DP, the sensitivity analysis of local updates, and generalization analysis. At last, we empirically confirm that our algorithms achieve state-of-the-art (SOTA) performance compared with existing SOTA baselines in DPFL.