DeepSeek-V3 Technical Report

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. The model checkpoints are available at https://github.com/deepseek-ai/DeepSeek-V3.

When Can Proxies Improve the Sample Complexity of Preference Learning?

We address the problem of reward hacking, where maximising a proxy reward does not necessarily increase the true reward. This is a key concern for Large Language Models (LLMs), as they are often fine-tuned on human preferences that may not accurately reflect a true objective. Existing work uses various tricks such as regularisation, tweaks to the reward model, and reward hacking detectors, to limit the influence that such proxy preferences have on a model. Luckily, in many contexts such as medicine, education, and law, a sparse amount of expert data is often available. In these cases, it is often unclear whether the addition of proxy data can improve policy learning. We outline a set of sufficient conditions on proxy feedback that, if satisfied, indicate that proxy data can provably improve the sample complexity of learning the ground truth policy. These conditions can inform the data collection process for specific tasks. The result implies a parameterisation for LLMs that achieves this improved sample complexity. We detail how one can adapt existing architectures to yield this improved sample complexity.

Generative modeling assisted simulation of measurement-altered quantum criticality

In quantum many-body systems, measurements can induce qualitative new features, but their simulation is hindered by the exponential complexity involved in sampling the measurement results. We propose to use machine learning to assist the simulation of measurement-induced quantum phenomena. In particular, we focus on the measurement-altered quantum criticality protocol and generate local reduced density matrices of the critical chain given random measurement results. Such generation is enabled by a physics-preserving conditional diffusion generative model, which learns an observation-indexed probability distribution of an ensemble of quantum states, and then samples from that distribution given an observation.

Deep Learning Algorithms for Mean Field Optimal Stopping in Finite Space and Discrete Time

Optimal stopping is a fundamental problem in optimization that has found applications in risk management, finance, economics, and recently in the fields of computer science. We extend the standard framework to a multi-agent setting, named multi-agent optimal stopping (MAOS), where a group of agents cooperatively solves finite-space, discrete-time optimal stopping problems. Solving the finite-agent case is computationally prohibitive when the number of agents is very large, so this work studies the mean field optimal stopping (MFOS) problem, obtained as the number of agents approaches infinity. We prove that MFOS provides a good approximate solution to MAOS. We also prove a dynamic programming principle (DPP), based on the theory of mean field control. We then propose two deep learning methods: one simulates full trajectories to learn optimal decisions, whereas the other leverages DPP with backward induction; both methods train neural networks for the optimal stopping decisions. We demonstrate the effectiveness of these approaches through numerical experiments on 6 different problems in spatial dimension up to 300. To the best of our knowledge, this is the first work to study MFOS in finite space and discrete time, and to propose efficient and scalable computational methods for this type of problem.

Plug-and-Play Controllable Generation for Discrete Masked Models

This article makes discrete masked models for the generative modeling of discrete data controllable. The goal is to generate samples of a discrete random variable that adheres to a posterior distribution, satisfies specific constraints, or optimizes a reward function. This methodological development enables broad applications across downstream tasks such as class-specific image generation and protein design. Existing approaches for controllable generation of masked models typically rely on task-specific fine-tuning or additional modifications, which can be inefficient and resource-intensive. To overcome these limitations, we propose a novel plug-and-play framework based on importance sampling that bypasses the need for training a conditional score. Our framework is agnostic to the choice of control criteria, requires no gradient information, and is well-suited for tasks such as posterior sampling, Bayesian inverse problems, and constrained generation. We demonstrate the effectiveness of our approach through extensive experiments, showcasing its versatility across multiple domains, including protein design.

Plain-Det: A Plain Multi-Dataset Object Detector

Recent advancements in large-scale foundational models have sparked widespread interest in training highly proficient large vision models. A common consensus revolves around the necessity of aggregating extensive, high-quality annotated data. However, given the inherent challenges in annotating dense tasks in computer vision, such as object detection and segmentation, a practical strategy is to combine and leverage all available data for training purposes. In this work, we propose Plain-Det, which offers flexibility to accommodate new datasets, robustness in performance across diverse datasets, training efficiency, and compatibility with various detection architectures. We utilize Def-DETR, with the assistance of Plain-Det, to achieve a mAP of 51.9 on COCO, matching the current state-of-the-art detectors. We conduct extensive experiments on 13 downstream datasets and Plain-Det demonstrates strong generalization capability. Code is release at https://github.com/ChengShiest/Plain-Det

Token-Mol 1.0: Tokenized drug design with large language model

Significant interests have recently risen in leveraging sequence-based large language models (LLMs) for drug design. However, most current applications of LLMs in drug discovery lack the ability to comprehend three-dimensional (3D) structures, thereby limiting their effectiveness in tasks that explicitly involve molecular conformations. In this study, we introduced Token-Mol, a token-only 3D drug design model. This model encodes all molecular information, including 2D and 3D structures, as well as molecular property data, into tokens, which transforms classification and regression tasks in drug discovery into probabilistic prediction problems, thereby enabling learning through a unified paradigm. Token-Mol is built on the transformer decoder architecture and trained using random causal masking techniques. Additionally, we proposed the Gaussian cross-entropy (GCE) loss function to overcome the challenges in regression tasks, significantly enhancing the capacity of LLMs to learn continuous numerical values. Through a combination of fine-tuning and reinforcement learning (RL), Token-Mol achieves performance comparable to or surpassing existing task-specific methods across various downstream tasks, including pocket-based molecular generation, conformation generation, and molecular property prediction. Compared to existing molecular pre-trained models, Token-Mol exhibits superior proficiency in handling a wider range of downstream tasks essential for drug design. Notably, our approach improves regression task accuracy by approximately 30% compared to similar token-only methods. Token-Mol overcomes the precision limitations of token-only models and has the potential to integrate seamlessly with general models such as ChatGPT, paving the way for the development of a universal artificial intelligence drug design model that facilitates rapid and high-quality drug design by experts.

Human-level molecular optimization driven by mol-gene evolution

De novo molecule generation allows the search for more drug-like hits across a vast chemical space. However, lead optimization is still required, and the process of optimizing molecular structures faces the challenge of balancing structural novelty with pharmacological properties. This study introduces the Deep Genetic Molecular Modification Algorithm (DGMM), which brings structure modification to the level of medicinal chemists. A discrete variational autoencoder (D-VAE) is used in DGMM to encode molecules as quantization code, mol-gene, which incorporates deep learning into genetic algorithms for flexible structural optimization. The mol-gene allows for the discovery of pharmacologically similar but structurally distinct compounds, and reveals the trade-offs of structural optimization in drug discovery. We demonstrate the effectiveness of the DGMM in several applications.

Structured Learning of Compositional Sequential Interventions

We consider sequential treatment regimes where each unit is exposed to combinations of interventions over time. When interventions are described by qualitative labels, such as ``close schools for a month due to a pandemic'' or ``promote this podcast to this user during this week'', it is unclear which appropriate structural assumptions allow us to generalize behavioral predictions to previously unseen combinatorial sequences. Standard black-box approaches mapping sequences of categorical variables to outputs are applicable, but they rely on poorly understood assumptions on how reliable generalization can be obtained, and may underperform under sparse sequences, temporal variability, and large action spaces. To approach that, we pose an explicit model for \emph{composition}, that is, how the effect of sequential interventions can be isolated into modules, clarifying which data conditions allow for the identification of their combined effect at different units and time steps. We show the identification properties of our compositional model, inspired by advances in causal matrix factorization methods but focusing on predictive models for novel compositions of interventions instead of matrix completion tasks and causal effect estimation. We compare our approach to flexible but generic black-box models to illustrate how structure aids prediction in sparse data conditions.

Trivialized Momentum Facilitates Diffusion Generative Modeling on Lie Groups

The generative modeling of data on manifold is an important task, for which diffusion models in flat spaces typically need nontrivial adaptations. This article demonstrates how a technique called `trivialization' can transfer the effectiveness of diffusion models in Euclidean spaces to Lie groups. In particular, an auxiliary momentum variable was algorithmically introduced to help transport the position variable between data distribution and a fixed, easy-to-sample distribution. Normally, this would incur further difficulty for manifold data because momentum lives in a space that changes with the position. However, our trivialization technique creates to a new momentum variable that stays in a simple $\textbf{fixed vector space}$. This design, together with a manifold preserving integrator, simplifies implementation and avoids inaccuracies created by approximations such as projections to tangent space and manifold, which were typically used in prior work, hence facilitating generation with high-fidelity and efficiency. The resulting method achieves state-of-the-art performance on protein and RNA torsion angle generation and sophisticated torus datasets. We also, arguably for the first time, tackle the generation of data on high-dimensional Special Orthogonal and Unitary groups, the latter essential for quantum problems.