Chen Dun

ByteDance Seed, Jiaze Chen, Tiantian Fan et al. · bytedance

We introduce Seed1.5-Thinking, capable of reasoning through thinking before responding, resulting in improved performance on a wide range of benchmarks. Seed1.5-Thinking achieves 86.7 on AIME 2024, 55.0 on Codeforces and 77.3 on GPQA, demonstrating excellent reasoning abilities in STEM and coding. Beyond reasoning tasks, the method demonstrates notable generalization across diverse domains. For instance, it surpasses DeepSeek R1 by 8% in win rate on non-reasoning tasks, indicating its broader applicability. Compared to other state-of-the-art reasoning models, Seed1.5-Thinking is a Mixture-of-Experts (MoE) model with a relatively small size, featuring 20B activated and 200B total parameters. As part of our effort to assess generalized reasoning, we develop two internal benchmarks, BeyondAIME and Codeforces, both of which will be publicly released to support future research. Model trial link: https://www.volcengine.com/experience/ark.

Chen Dun, Mirian Hipolito, Chris Jermaine et al.

Asynchronous learning protocols have regained attention lately, especially in the Federated Learning (FL) setup, where slower clients can severely impede the learning process. Herein, we propose \texttt{AsyncDrop}, a novel asynchronous FL framework that utilizes dropout regularization to handle device heterogeneity in distributed settings. Overall, \texttt{AsyncDrop} achieves better performance compared to state of the art asynchronous methodologies, while resulting in less communication and training time overheads. The key idea revolves around creating ``submodels'' out of the global model, and distributing their training to workers, based on device heterogeneity. We rigorously justify that such an approach can be theoretically characterized. We implement our approach and compare it against other asynchronous baselines, both by design and by adapting existing synchronous FL algorithms to asynchronous scenarios. Empirically, \texttt{AsyncDrop} reduces the communication cost and training time, while matching or improving the final test accuracy in diverse non-i.i.d. FL scenarios.

Qihan Wang, Chen Dun, Fangshuo Liao et al.

Recent work on the Lottery Ticket Hypothesis (LTH) shows that there exist ``\textit{winning tickets}'' in large neural networks. These tickets represent ``sparse'' versions of the full model that can be trained independently to achieve comparable accuracy with respect to the full model. However, finding the winning tickets requires one to \emph{pretrain} the large model for at least a number of epochs, which can be a burdensome task, especially when the original neural network gets larger. In this paper, we explore how one can efficiently identify the emergence of such winning tickets, and use this observation to design efficient pretraining algorithms. For clarity of exposition, our focus is on convolutional neural networks (CNNs). To identify good filters, we propose a novel filter distance metric that well-represents the model convergence. As our theory dictates, our filter analysis behaves consistently with recent findings of neural network learning dynamics. Motivated by these observations, we present the \emph{LOttery ticket through Filter-wise Training} algorithm, dubbed as \textsc{LoFT}. \textsc{LoFT} is a model-parallel pretraining algorithm that partitions convolutional layers by filters to train them independently in a distributed setting, resulting in reduced memory and communication costs during pretraining. Experiments show that \textsc{LoFT} $i)$ preserves and finds good lottery tickets, while $ii)$ it achieves non-trivial computation and communication savings, and maintains comparable or even better accuracy than other pretraining methods.

Chen Dun, Mirian Hipolito Garcia, Guoqing Zheng et al.

Large Language Models (LLMs) have the ability to solve a variety of tasks, such as text summarization and mathematical questions, just out of the box, but they are often trained with a single task in mind. Due to high computational costs, the current trend is to use prompt instruction tuning to better adjust monolithic, pretrained LLMs for new -- but often individual -- downstream tasks. Thus, how one would expand prompt tuning to handle -- concomitantly -- heterogeneous tasks and data distributions is a widely open question. To address this gap, we suggest the use of \emph{Mixture of Prompts}, or MoPs, associated with smart gating functionality: the latter -- whose design is one of the contributions of this paper -- can identify relevant skills embedded in different groups of prompts and dynamically assign combined experts (i.e., collection of prompts), based on the target task. Additionally, MoPs are empirically agnostic to any model compression technique applied -- for efficiency reasons -- as well as instruction data source and task composition. In practice, MoPs can simultaneously mitigate prompt training "interference" in multi-task, multi-source scenarios (e.g., task and data heterogeneity across sources), as well as possible implications from model approximations. As a highlight, MoPs manage to decrease final perplexity from $\sim20\%$ up to $\sim70\%$, as compared to baselines, in the federated scenario, and from $\sim 3\%$ up to $\sim30\%$ in the centralized scenario.

John Chen, Chen Dun, Anastasios Kyrillidis

Advances in Semi-Supervised Learning (SSL) have almost entirely closed the gap between SSL and Supervised Learning at a fraction of the number of labels. However, recent performance improvements have often come \textit{at the cost of significantly increased training computation}. To address this, we propose Curriculum Batch Size (CBS), \textit{an unlabeled batch size curriculum which exploits the natural training dynamics of deep neural networks.} A small unlabeled batch size is used in the beginning of training and is gradually increased to the end of training. A fixed curriculum is used regardless of dataset, model or number of epochs, and reduced training computations is demonstrated on all settings. We apply CBS, strong labeled augmentation, Curriculum Pseudo Labeling (CPL) \citep{FlexMatch} to FixMatch \citep{FixMatch} and term the new SSL algorithm Fast FixMatch. We perform an ablation study to show that strong labeled augmentation and/or CPL do not significantly reduce training computations, but, in synergy with CBS, they achieve optimal performance. Fast FixMatch also achieves substantially higher data utilization compared to previous state-of-the-art. Fast FixMatch achieves between $2.1\times$ - $3.4\times$ reduced training computations on CIFAR-10 with all but 40, 250 and 4000 labels removed, compared to vanilla FixMatch, while attaining the same cited state-of-the-art error rate \citep{FixMatch}. Similar results are achieved for CIFAR-100, SVHN and STL-10. Finally, Fast MixMatch achieves between $2.6\times$ - $3.3\times$ reduced training computations in federated SSL tasks and online/streaming learning SSL tasks, which further demonstrate the generializbility of Fast MixMatch to different scenarios and tasks.

Yehya Farhat, Hamza ElMokhtar Shili, Fangshuo Liao et al.

Mixture-of-Experts (MoEs) achieve scalability by dynamically activating subsets of their components. Yet, understanding how expertise emerges through joint training of gating mechanisms and experts remains incomplete, especially in scenarios without clear task partitions. Motivated by inference costs and data heterogeneity, we study how joint training of gating functions and experts can dynamically allocate domain-specific expertise across multiple underlying data distributions. As an outcome of our framework, we develop an instance tailored specifically to decentralized training scenarios, introducing \textit{Dynamically Decentralized Orchestration of MoEs} or \texttt{DDOME}. \texttt{DDOME} leverages heterogeneity emerging from distributional shifts across decentralized data sources to specialize experts dynamically. By integrating a pretrained common expert to inform a gating function, \texttt{DDOME} achieves personalized expert subset selection on-the-fly, facilitating just-in-time personalization. We empirically validate \texttt{DDOME} within a Federated Learning (FL) context: \texttt{DDOME} attains from 4\% up to an 24\% accuracy improvement over state-of-the-art FL baselines in image and text classification tasks, while maintaining competitive zero-shot generalization capabilities. Furthermore, we provide theoretical insights confirming that the joint gating-experts training is critical for achieving meaningful expert specialization.

Chen Dun, Qiutai Pan, Shikai Jin et al.

Determining the structure of a protein has been a decades-long open question. A protein's three-dimensional structure often poses nontrivial computation costs, when classical simulation algorithms are utilized. Advances in the transformer neural network architecture -- such as AlphaFold2 -- achieve significant improvements for this problem, by learning from a large dataset of sequence information and corresponding protein structures. Yet, such methods only focus on sequence information; other available prior knowledge, such as protein crystallography and partial structure of amino acids, could be potentially utilized. To the best of our knowledge, we propose the first transformer-based model that directly utilizes protein crystallography and partial structure information to predict the electron density maps of proteins. Via two new datasets of peptide fragments (2-residue and 15-residue) , we demonstrate our method, dubbed \texttt{CrysFormer}, can achieve accurate predictions, based on a much smaller dataset size and with reduced computation costs.

Junting Zhou, Tingjia Miao, Yiyan Liao et al.

Advancement in Large Language Models (LLMs) reasoning capabilities enables them to solve scientific problems with enhanced efficacy. Thereby, a high-quality benchmark for comprehensive and appropriate assessment holds significance, while existing ones either confront the risk of data contamination or lack involved disciplines. To be specific, due to the data source overlap of LLMs training and static benchmark, the keys or number pattern of answers inadvertently memorized (i.e. data contamination), leading to systematic overestimation of their reasoning capabilities, especially numerical reasoning. We propose SciDA, a multidisciplinary benchmark that consists exclusively of over 1k Olympic-level numerical computation problems, allowing randomized numerical initializations for each inference round to avoid reliance on fixed numerical patterns. We conduct a series of experiments with both closed-source and open-source top-performing LLMs, and it is observed that the performance of LLMs drop significantly under random numerical initialization. Thus, we provide truthful and unbiased assessments of the numerical reasoning capabilities of LLMs. The data is available at https://huggingface.co/datasets/m-a-p/SciDA

Haoming Wang, Haoyang Zou, Huatong Song et al. · pku

The development of autonomous agents for graphical user interfaces (GUIs) presents major challenges in artificial intelligence. While recent advances in native agent models have shown promise by unifying perception, reasoning, action, and memory through end-to-end learning, open problems remain in data scalability, multi-turn reinforcement learning (RL), the limitations of GUI-only operation, and environment stability. In this technical report, we present UI-TARS-2, a native GUI-centered agent model that addresses these challenges through a systematic training methodology: a data flywheel for scalable data generation, a stabilized multi-turn RL framework, a hybrid GUI environment that integrates file systems and terminals, and a unified sandbox platform for large-scale rollouts. Empirical evaluation demonstrates that UI-TARS-2 achieves significant improvements over its predecessor UI-TARS-1.5. On GUI benchmarks, it reaches 88.2 on Online-Mind2Web, 47.5 on OSWorld, 50.6 on WindowsAgentArena, and 73.3 on AndroidWorld, outperforming strong baselines such as Claude and OpenAI agents. In game environments, it attains a mean normalized score of 59.8 across a 15-game suite-roughly 60% of human-level performance-and remains competitive with frontier proprietary models (e.g., OpenAI o3) on LMGame-Bench. Additionally, the model can generalize to long-horizon information-seeking tasks and software engineering benchmarks, highlighting its robustness across diverse agent tasks. Detailed analyses of training dynamics further provide insights into achieving stability and efficiency in large-scale agent RL. These results underscore UI-TARS-2's potential to advance the state of GUI agents and exhibit strong generalization to real-world interactive scenarios.

Kaixuan Ji, Guanlin Liu, Ning Dai et al.

Reinforcement Learning (RL) plays a crucial role in aligning large language models (LLMs) with human preferences and improving their ability to perform complex tasks. However, current approaches either require significant computational resources due to the use of multiple models and extensive online sampling for training (e.g., PPO) or are framed as bandit problems (e.g., DPO, DRO), which often struggle with multi-step reasoning tasks, such as math problem solving and complex reasoning that involve long chains of thought. To overcome these limitations, we introduce Direct Q-function Optimization (DQO), which formulates the response generation process as a Markov Decision Process (MDP) and utilizes the soft actor-critic (SAC) framework to optimize a Q-function directly parameterized by the language model. The MDP formulation of DQO offers structural advantages over bandit-based methods, enabling more effective process supervision. Experimental results on two math problem-solving datasets, GSM8K and MATH, demonstrate that DQO outperforms previous methods, establishing it as a promising offline reinforcement learning approach for aligning language models.

Ning Dai, Zheng Wu, Renjie Zheng et al.

Reinforcement learning (RL) with unit test feedback has enhanced large language models' (LLMs) code generation, but relies on sparse rewards provided only after complete code evaluation, limiting learning efficiency and incremental improvements. When generated code fails all unit tests, no learning signal is received, hindering progress on complex tasks. To address this, we propose a Process Reward Model (PRM) that delivers dense, line-level feedback on code correctness during generation, mimicking human code refinement and providing immediate guidance. We explore various strategies for training PRMs and integrating them into the RL framework, finding that using PRMs both as dense rewards and for value function initialization significantly boosts performance. Our experimental results also highlight the effectiveness of PRMs in enhancing RL-driven code generation, especially for long-horizon scenarios.

Weizhe Chen, Zhicheng Zhang, Guanlin Liu et al.

Since the release of ChatGPT, large language models (LLMs) have demonstrated remarkable capabilities across various domains. A key challenge in developing these general capabilities is efficiently sourcing diverse, high-quality data. This becomes especially critical in reasoning-related tasks with sandbox checkers, such as math or code, where the goal is to generate correct solutions to specific problems with higher probability. In this work, we introduce Flaming-hot Initiation with Regular Execution (FIRE) sampling, a simple yet highly effective method to efficiently find good responses. Our empirical findings show that FIRE sampling enhances inference-time generation quality and also benefits training in the alignment stage. Furthermore, we explore how FIRE sampling improves performance by promoting diversity and analyze the impact of employing FIRE at different positions within a response.

Zhenwei Dai, Chen Dun, Yuxin Tang et al.

Federated learning enables many local devices to train a deep learning model jointly without sharing the local data. Currently, most of federated training schemes learns a global model by averaging the parameters of local models. However, most of these training schemes suffer from high communication cost resulted from transmitting full local model parameters. Moreover, directly averaging model parameters leads to a significant performance degradation, due to the class-imbalanced non-iid data on different devices. Especially for the real life federated learning tasks involving extreme classification, (1) communication becomes the main bottleneck since the model size increases proportionally to the number of output classes; (2) extreme classification (such as user recommendation) normally have extremely imbalanced classes and heterogeneous data on different devices. To overcome this problem, we propose federated multiple label hashing (FedMLH), which leverages label hashing to simultaneously reduce the model size (up to 3.40X decrease) with communication cost (up to 18.75X decrease) and achieves significant better accuracy (up to 35.5%} relative accuracy improvement) and faster convergence rate (up to 5.5X increase) for free on the federated extreme classification tasks compared to federated average algorithm.

Chen Dun, Cameron R. Wolfe, Christopher M. Jermaine et al.

We propose ResIST, a novel distributed training protocol for Residual Networks (ResNets). ResIST randomly decomposes a global ResNet into several shallow sub-ResNets that are trained independently in a distributed manner for several local iterations, before having their updates synchronized and aggregated into the global model. In the next round, new sub-ResNets are randomly generated and the process repeats until convergence. By construction, per iteration, ResIST communicates only a small portion of network parameters to each machine and never uses the full model during training. Thus, ResIST reduces the per-iteration communication, memory, and time requirements of ResNet training to only a fraction of the requirements of full-model training. In comparison to common protocols, like data-parallel training and data-parallel training with local SGD, ResIST yields a decrease in communication and compute requirements, while being competitive with respect to model performance.

Cameron R. Wolfe, Jingkang Yang, Arindam Chowdhury et al.

The graph convolutional network (GCN) is a go-to solution for machine learning on graphs, but its training is notoriously difficult to scale both in terms of graph size and the number of model parameters. Although some work has explored training on large-scale graphs (e.g., GraphSAGE, ClusterGCN, etc.), we pioneer efficient training of large-scale GCN models (i.e., ultra-wide, overparameterized models) with the proposal of a novel, distributed training framework. Our proposed training methodology, called GIST, disjointly partitions the parameters of a GCN model into several, smaller sub-GCNs that are trained independently and in parallel. In addition to being compatible with all GCN architectures and existing sampling techniques for efficient GCN training, GIST i) improves model performance, ii) scales to training on arbitrarily large graphs, iii) decreases wall-clock training time, and iv) enables the training of markedly overparameterized GCN models. Remarkably, with GIST, we train an astonishgly-wide 32,768-dimensional GraphSAGE model, which exceeds the capacity of a single GPU by a factor of 8x, to SOTA performance on the Amazon2M dataset.

Binhang Yuan, Cameron R. Wolfe, Chen Dun et al.

Distributed machine learning (ML) can bring more computational resources to bear than single-machine learning, thus enabling reductions in training time. Distributed learning partitions models and data over many machines, allowing model and dataset sizes beyond the available compute power and memory of a single machine. In practice though, distributed ML is challenging when distribution is mandatory, rather than chosen by the practitioner. In such scenarios, data could unavoidably be separated among workers due to limited memory capacity per worker or even because of data privacy issues. There, existing distributed methods will utterly fail due to dominant transfer costs across workers, or do not even apply. We propose a new approach to distributed fully connected neural network learning, called independent subnet training (IST), to handle these cases. In IST, the original network is decomposed into a set of narrow subnetworks with the same depth. These subnetworks are then trained locally before parameters are exchanged to produce new subnets and the training cycle repeats. Such a naturally "model parallel" approach limits memory usage by storing only a portion of network parameters on each device. Additionally, no requirements exist for sharing data between workers (i.e., subnet training is local and independent) and communication volume and frequency are reduced by decomposing the original network into independent subnets. These properties of IST can cope with issues due to distributed data, slow interconnects, or limited device memory, making IST a suitable approach for cases of mandatory distribution. We show experimentally that IST results in training times that are much lower than common distributed learning approaches.

Jiechuan Jiang, Chen Dun, Tiejun Huang et al.

Learning to cooperate is crucially important in multi-agent environments. The key is to understand the mutual interplay between agents. However, multi-agent environments are highly dynamic, where agents keep moving and their neighbors change quickly. This makes it hard to learn abstract representations of mutual interplay between agents. To tackle these difficulties, we propose graph convolutional reinforcement learning, where graph convolution adapts to the dynamics of the underlying graph of the multi-agent environment, and relation kernels capture the interplay between agents by their relation representations. Latent features produced by convolutional layers from gradually increased receptive fields are exploited to learn cooperation, and cooperation is further improved by temporal relation regularization for consistency. Empirically, we show that our method substantially outperforms existing methods in a variety of cooperative scenarios.