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.

While speculative decoding has recently appeared as a promising direction for accelerating the inference of large language models (LLMs), the speedup and scalability are strongly bounded by the token acceptance rate. Prevalent methods usually organize predicted tokens as independent chains or fixed token trees, which fails to generalize to diverse query distributions. In this paper, we propose DySpec, a faster speculative decoding algorithm with a novel dynamic token tree structure. We begin by bridging the draft distribution and acceptance rate from intuitive and empirical clues, and successfully show that the two variables are strongly correlated. Based on this, we employ a greedy strategy to dynamically expand the token tree at run time. Theoretically, we show that our method can achieve optimal results under mild assumptions. Empirically, DySpec yields a higher acceptance rate and speedup than fixed trees. DySpec can drastically improve the throughput and reduce the latency of token generation across various data distribution and model sizes, which significantly outperforms strong competitors, including Specinfer and Sequoia. Under low temperature setting, DySpec can improve the throughput up to 9.1$\times$ and reduce the latency up to 9.4$\times$ on Llama2-70B. Under high temperature setting, DySpec can also improve the throughput up to 6.21$\times$, despite the increasing difficulty of speculating more than one token per step for draft model.

Simulated Patients (SPs) play a crucial role in clinical medical education by providing realistic scenarios for student practice. However, the high cost of training and hiring qualified SPs, along with the heavy workload and potential risks they face in consistently portraying actual patients, limit students' access to this type of clinical training. Consequently, the integration of computer program-based simulated patients has emerged as a valuable educational tool in recent years. With the rapid development of Large Language Models (LLMs), their exceptional capabilities in conversational artificial intelligence and role-playing have been demonstrated, making them a feasible option for implementing Virtual Simulated Patient (VSP). In this paper, we present an integrated model-agnostic framework called CureFun that harnesses the potential of LLMs in clinical medical education. This framework facilitates natural conversations between students and simulated patients, evaluates their dialogue, and provides suggestions to enhance students' clinical inquiry skills. Through comprehensive evaluations, our approach demonstrates more authentic and professional SP-scenario dialogue flows compared to other LLM-based chatbots, thus proving its proficiency in simulating patients. Additionally, leveraging CureFun's evaluation ability, we assess several medical LLMs and discuss the possibilities and limitations of using LLMs as virtual doctors from the perspective of their diagnostic abilities.

Transformer-based Large Language Models (LLMs) have made a significant impact on various domains. However, LLMs' efficiency suffers from both heavy computation and memory overheads. Compression techniques like sparsification and quantization are commonly used to mitigate the gap between LLM's computation/memory overheads and hardware capacity. However, existing GPU and transformer-based accelerators cannot efficiently process compressed LLMs, due to the following unresolved challenges: low computational efficiency, underutilized memory bandwidth, and large compilation overheads. This paper proposes FlightLLM, enabling efficient LLMs inference with a complete mapping flow on FPGAs. In FlightLLM, we highlight an innovative solution that the computation and memory overhead of LLMs can be solved by utilizing FPGA-specific resources (e.g., DSP48 and heterogeneous memory hierarchy). We propose a configurable sparse DSP chain to support different sparsity patterns with high computation efficiency. Second, we propose an always-on-chip decode scheme to boost memory bandwidth with mixed-precision support. Finally, to make FlightLLM available for real-world LLMs, we propose a length adaptive compilation method to reduce the compilation overhead. Implemented on the Xilinx Alveo U280 FPGA, FlightLLM achieves 6.0$\times$ higher energy efficiency and 1.8$\times$ better cost efficiency against commercial GPUs (e.g., NVIDIA V100S) on modern LLMs (e.g., LLaMA2-7B) using vLLM and SmoothQuant under the batch size of one. FlightLLM beats NVIDIA A100 GPU with 1.2$\times$ higher throughput using the latest Versal VHK158 FPGA.

The high-accuracy and resource-intensive deep neural networks (DNNs) have been widely adopted by live video analytics (VA), where camera videos are streamed over the network to resource-rich edge/cloud servers for DNN inference. Common video encoding configurations (e.g., resolution and frame rate) have been identified with significant impacts on striking the balance between bandwidth consumption and inference accuracy and therefore their adaption scheme has been a focus of optimization. However, previous profiling-based solutions suffer from high profiling cost, while existing deep reinforcement learning (DRL) based solutions may achieve poor performance due to the usage of fixed reward function for training the agent, which fails to craft the application goals in various scenarios. In this paper, we propose ILCAS, the first imitation learning (IL) based configuration-adaptive VA streaming system. Unlike DRL-based solutions, ILCAS trains the agent with demonstrations collected from the expert which is designed as an offline optimal policy that solves the configuration adaption problem through dynamic programming. To tackle the challenge of video content dynamics, ILCAS derives motion feature maps based on motion vectors which allow ILCAS to visually ``perceive'' video content changes. Moreover, ILCAS incorporates a cross-camera collaboration scheme to exploit the spatio-temporal correlations of cameras for more proper configuration selection. Extensive experiments confirm the superiority of ILCAS compared with state-of-the-art solutions, with 2-20.9% improvement of mean accuracy and 19.9-85.3% reduction of chunk upload lag.

Prevalent supervised learning methods in natural language processing (NLP) are notoriously data-hungry, which demand large amounts of high-quality annotated data. In practice, acquiring such data is a costly endeavor. Recently, the superior few-shot performance of large language models (LLMs) has propelled the development of dataset generation, where the training data are solely synthesized from LLMs. However, such an approach usually suffers from low-quality issues, and requires orders of magnitude more labeled data to achieve satisfactory performance. To fully exploit the potential of LLMs and make use of massive unlabeled data, we propose LLMaAA, which takes LLMs as annotators and puts them into an active learning loop to determine what to annotate efficiently. To learn robustly with pseudo labels, we optimize both the annotation and training processes: (1) we draw k-NN examples from a small demonstration pool as in-context examples, and (2) we adopt the example reweighting technique to assign training samples with learnable weights. Compared with previous approaches, LLMaAA features both efficiency and reliability. We conduct experiments and analysis on two classic NLP tasks, named entity recognition and relation extraction. With LLMaAA, task-specific models trained from LLM-generated labels can outperform the teacher within only hundreds of annotated examples, which is much more cost-effective than other baselines.

Large Language Models (LLMs) can acquire extensive world knowledge through pre-training on large corpora. However, due to exposure to low-quality data, LLMs may exhibit harmful behavior without aligning with human values. The dominant approach for steering LLMs towards beneficial behavior involves Reinforcement Learning with Human Feedback (RLHF), with Proximal Policy Optimization (PPO) serving as the default RL optimizer. Despite its effectiveness, PPO has limitations when optimizing rewards trained from comparison-based loss. Primarily, PPO is not invariant to equivalent reward functions containing identical preference information due to the need to calibrate the reward scale. Additionally, PPO's necessity for token-wise updates introduces complexity in both function approximation and algorithm design compared to trajectory-wise optimization. This paper proposes a new framework, reinforcement learning with relative feedback, and a novel trajectory-wise policy gradient algorithm, Pairwise Proximal Policy Optimization (P3O) that operates directly on comparative rewards. We show theoretically that P3O is invariant to equivalent rewards and avoids the complexity of PPO. Empirical evaluations demonstrate that P3O outperforms PPO in the KL-Reward trade-off and can align with human preferences as well as or better than prior methods. In summary, this work introduces a simpler yet effective approach for aligning LLMs to human preferences through relative feedback.

Incorporating multiple knowledge sources is proven to be beneficial for answering complex factoid questions. To utilize multiple knowledge bases (KB), previous works merge all KBs into a single graph via entity alignment and reduce the problem to question-answering (QA) over the fused KB. In reality, various link relations between KBs might be adopted in QA over multi-KBs. In addition to the identity between the alignable entities (i.e. full link), unalignable entities expressing the different aspects or types of an abstract concept may also be treated identical in a question (i.e. partial link). Hence, the KB fusion in prior works fails to represent all types of links, restricting their ability to comprehend multi-KBs for QA. In this work, we formulate the novel Multi-KB-QA task that leverages the full and partial links among multiple KBs to derive correct answers, a benchmark with diversified link and query types is also constructed to efficiently evaluate Multi-KB-QA performance. Finally, we propose a method for Multi-KB-QA that encodes all link relations in the KB embedding to score and rank candidate answers. Experiments show that our method markedly surpasses conventional KB-QA systems in Multi-KB-QA, justifying the necessity of devising this task.

Image-text retrieval in remote sensing aims to provide flexible information for data analysis and application. In recent years, state-of-the-art methods are dedicated to ``scale decoupling'' and ``semantic decoupling'' strategies to further enhance the capability of representation. However, these previous approaches focus on either the disentangling scale or semantics but ignore merging these two ideas in a union model, which extremely limits the performance of cross-modal retrieval models. To address these issues, we propose a novel Scale-Semantic Joint Decoupling Network (SSJDN) for remote sensing image-text retrieval. Specifically, we design the Bidirectional Scale Decoupling (BSD) module, which exploits Salience Feature Extraction (SFE) and Salience-Guided Suppression (SGS) units to adaptively extract potential features and suppress cumbersome features at other scales in a bidirectional pattern to yield different scale clues. Besides, we design the Label-supervised Semantic Decoupling (LSD) module by leveraging the category semantic labels as prior knowledge to supervise images and texts probing significant semantic-related information. Finally, we design a Semantic-guided Triple Loss (STL), which adaptively generates a constant to adjust the loss function to improve the probability of matching the same semantic image and text and shorten the convergence time of the retrieval model. Our proposed SSJDN outperforms state-of-the-art approaches in numerical experiments conducted on four benchmark remote sensing datasets.