Ya Wang

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.

Lianghui Zhu, Yuxin Fang, Bencheng Liao et al.

Scaling depth is a key driver for large language models (LLMs). Yet, as LLMs become deeper, they often suffer from signal degradation: informative features formed in shallow layers are gradually diluted by repeated residual updates, making them harder to recover in deeper layers. We introduce mixture-of-depths attention (MoDA), a mechanism that allows each attention head to attend to sequence KV pairs at the current layer and depth KV pairs from preceding layers. We further describe a hardware-efficient algorithm for MoDA that resolves non-contiguous memory-access patterns, achieving 97.3% of FlashAttention-2's efficiency at a sequence length of 64K. Experiments on 1.5B-parameter models demonstrate that MoDA consistently outperforms strong baselines. Notably, it improves average perplexity by 0.2 across 10 validation benchmarks and increases average performance by 2.11% on 10 downstream tasks, with a negligible 3.7% FLOPs computational overhead. We also find that combining MoDA with post-norm yields better performance than using it with pre-norm. These results suggest that MoDA is a promising primitive for depth scaling. Code is released at https://github.com/hustvl/MoDA .

Julian Wörmann, Daniel Bogdoll, Christian Brunner et al.

The availability of representative datasets is an essential prerequisite for many successful artificial intelligence and machine learning models. However, in real life applications these models often encounter scenarios that are inadequately represented in the data used for training. There are various reasons for the absence of sufficient data, ranging from time and cost constraints to ethical considerations. As a consequence, the reliable usage of these models, especially in safety-critical applications, is still a tremendous challenge. Leveraging additional, already existing sources of knowledge is key to overcome the limitations of purely data-driven approaches. Knowledge augmented machine learning approaches offer the possibility of compensating for deficiencies, errors, or ambiguities in the data, thus increasing the generalization capability of the applied models. Even more, predictions that conform with knowledge are crucial for making trustworthy and safe decisions even in underrepresented scenarios. This work provides an overview of existing techniques and methods in the literature that combine data-driven models with existing knowledge. The identified approaches are structured according to the categories knowledge integration, extraction and conformity. In particular, we address the application of the presented methods in the field of autonomous driving.

Chenguang Li, Jia Shi, Ya Wang et al.

In this paper, we propose an advanced approach in targeting the problem of monocular 3D lane detection by leveraging geometry structure underneath the process of 2D to 3D lane reconstruction. Inspired by previous methods, we first analyze the geometry heuristic between the 3D lane and its 2D representation on the ground and propose to impose explicit supervision based on the structure prior, which makes it achievable to build inter-lane and intra-lane relationships to facilitate the reconstruction of 3D lanes from local to global. Second, to reduce the structure loss in 2D lane representation, we directly extract BEV lane information from front view images, which tremendously eases the confusion of distant lane features in previous methods. Furthermore, we propose a novel task-specific data augmentation method by synthesizing new training data for both segmentation and reconstruction tasks in our pipeline, to counter the imbalanced data distribution of camera pose and ground slope to improve generalization on unseen data. Our work marks the first attempt to employ the geometry prior information into DNN-based 3D lane detection and makes it achievable for detecting lanes in an extra-long distance, doubling the original detection range. The proposed method can be smoothly adopted by other frameworks without extra costs. Experimental results show that our work outperforms state-of-the-art approaches by 3.8% F-Score on Apollo 3D synthetic dataset at real-time speed of 82 FPS without introducing extra parameters.

Zhijian Zhuo, Ya Wang, Yutao Zeng et al. · bytedance

Transformers have found extensive applications across various domains due to the powerful fitting capabilities. This success can be partially attributed to their inherent nonlinearity. Thus, in addition to the ReLU function employed in the original transformer architecture, researchers have explored alternative modules such as GeLU and SwishGLU to enhance nonlinearity and thereby augment representational capacity. In this paper, we propose a novel category of polynomial composition activations (PolyCom), designed to optimize the dynamics of transformers. Theoretically, we provide a comprehensive mathematical analysis of PolyCom, highlighting its enhanced expressivity and efficacy relative to other activation functions. Notably, we demonstrate that networks incorporating PolyCom achieve the $\textbf{optimal approximation rate}$, indicating that PolyCom networks require minimal parameters to approximate general smooth functions in Sobolev spaces. We conduct empirical experiments on the pre-training configurations of large language models (LLMs), including both dense and sparse architectures. By substituting conventional activation functions with PolyCom, we enable LLMs to capture higher-order interactions within the data, thus improving performance metrics in terms of accuracy and convergence rates. Extensive experimental results demonstrate the effectiveness of our method, showing substantial improvements over other activation functions. Code is available at https://github.com/BryceZhuo/PolyCom.

Zhijian Zhuo, Yutao Zeng, Ya Wang et al. · bytedance

Transformers have become the de facto architecture for a wide range of machine learning tasks, particularly in large language models (LLMs). Despite their remarkable performance, challenges remain in training deep transformer networks, especially regarding the position of layer normalization. While Pre-Norm structures facilitate more stable training owing to their stronger identity path, they often lead to suboptimal performance compared to Post-Norm. In this paper, we propose $\textbf{HybridNorm}$, a simple yet effective hybrid normalization strategy that integrates the advantages of both Pre-Norm and Post-Norm. Specifically, HybridNorm employs QKV normalization within the attention mechanism and Post-Norm in the feed-forward network (FFN) of each transformer block. We provide both theoretical insights and empirical evidence demonstrating that HybridNorm improves gradient flow and model robustness. Extensive experiments on large-scale transformer models, including both dense and sparse variants, show that HybridNorm consistently outperforms both Pre-Norm and Post-Norm approaches across multiple benchmarks. These findings highlight the potential of HybridNorm as a more stable and effective technique for improving the training and performance of deep transformer models. Code is available at https://github.com/BryceZhuo/HybridNorm.

Ya Wang, Zhijian Zhuo, Yutao Zeng et al. · bytedance

Training stability is a persistent challenge in the pre-training of large language models (LLMs), particularly for architectures such as Post-Norm Transformers, which are prone to gradient explosion and dissipation. In this paper, we propose Scale-Distribution Decoupling (SDD), a novel approach that stabilizes training by explicitly decoupling the scale and distribution of the weight matrix in fully-connected layers. SDD applies a normalization mechanism to regulate activations and a learnable scaling vector to maintain well-conditioned gradients, effectively preventing $\textbf{gradient explosion and dissipation}$. This separation improves optimization efficiency, particularly in deep networks, by ensuring stable gradient propagation. Experimental results demonstrate that our method stabilizes training across various LLM architectures and outperforms existing techniques in different normalization configurations. Furthermore, the proposed method is lightweight and compatible with existing frameworks, making it a practical solution for stabilizing LLM training. Code is available at https://github.com/kaihemo/SDD.

Tianye Li, Qi Liu, Hao Li et al.

Accurate global medium-range weather forecasting is fundamental to Earth system science. Most existing Transformer-based forecasting models adopt vision-centric architectures that neglect the Earth's spherical geometry and zonal periodicity. In addition, conventional autoregressive training is computationally expensive and limits forecast horizons due to error accumulation. To address these challenges, we propose the Shifted Earth Transformer (Searth Transformer), a physics-informed architecture that incorporates zonal periodicity and meridional boundaries into window-based self-attention for physically consistent global information exchange. We further introduce a Relay Autoregressive (RAR) fine-tuning strategy that enables learning long-range atmospheric evolution under constrained memory and computational budgets. Based on these methods, we develop YanTian, a global medium-range weather forecasting model. YanTian achieves higher accuracy than the high-resolution forecast of the European Centre for Medium-Range Weather Forecasts and performs competitively with state-of-the-art AI models at one-degree resolution, while requiring roughly 200 times lower computational cost than standard autoregressive fine-tuning. Furthermore, YanTian attains a longer skillful forecast lead time for Z500 (10.3 days) than HRES (9 days). Beyond weather forecasting, this work establishes a robust algorithmic foundation for predictive modeling of complex global-scale geophysical circulation systems, offering new pathways for Earth system science.

Hanwei Fan, Ya Wang, Sicheng Li et al.

With the continuous advancement of processors, modern micro-architecture designs have become increasingly complex. The vast design space presents significant challenges for human designers, making design space exploration (DSE) algorithms a significant tool for $μ$-arch design. In recent years, efforts have been made in the development of DSE algorithms, and promising results have been achieved. However, the existing DSE algorithms, e.g., Bayesian Optimization and ensemble learning, suffer from poor interpretability, hindering designers' understanding of the decision-making process. To address this limitation, we propose utilizing Fuzzy Neural Networks to induce and summarize knowledge and insights from the DSE process, enhancing interpretability and controllability. Furthermore, to improve efficiency, we introduce a multi-fidelity reinforcement learning approach, which primarily conducts exploration using cheap but less precise data, thereby substantially diminishing the reliance on costly data. Experimental results show that our method achieves excellent results with a very limited sample budget and successfully surpasses the current state-of-the-art. Our DSE framework is open-sourced and available at https://github.com/fanhanwei/FNN\_MFRL\_ArchDSE/\ .

Shengtian Mian, Ya Wang, Nannan Gu et al.

Windowed attention mechanisms were introduced to mitigate the issue of excessive computation inherent in global attention mechanisms. In this paper, we present FwNet-ECA, a novel method that utilizes Fourier transforms paired with learnable weight matrices to enhance the spectral features of images. This method establishes a global receptive field through Filter Enhancement and avoids the use of moving window attention. Additionally, we incorporate the Efficient Channel Attention (ECA) module to improve communication between different channels. Instead of relying on physically shifted windows, our approach leverages frequency domain enhancement to implicitly bridge information across spatial regions. We validate our model on the iCartoonFace dataset and conduct downstream tasks on ImageNet, demonstrating that our model achieves lower parameter counts and computational overheads compared to shifted window approaches, while maintaining competitive accuracy. Furthermore, our visualization operations clearly demonstrated that the Filter Enhancement technique achieves greater effectiveness in the model's shallow layers, where feature maps are relatively larger. This work offers a more efficient and effective alternative for leveraging attention mechanisms in visual processing tasks, alleviating the challenges associated with windowed attention models. Code is available at https://github.com/qingxiaoli/FwNet-ECA

Xin Huang, Ruibin Li, Tong Jia et al.

Vision-Language Models (VLMs) are essential for multimodal tasks, especially compositional reasoning (CR) tasks, which require distinguishing fine-grained semantic differences between visual and textual embeddings. However, existing methods primarily fine-tune the model by generating text-based hard negative samples, neglecting the importance of image-based negative samples, which results in insufficient training of the visual encoder and ultimately impacts the overall performance of the model. Moreover, negative samples are typically treated uniformly, without considering their difficulty levels, and the alignment of positive samples is insufficient, which leads to challenges in aligning difficult sample pairs. To address these issues, we propose Adaptive Hard Negative Perturbation Learning (AHNPL). AHNPL translates text-based hard negatives into the visual domain to generate semantically disturbed image-based negatives for training the model, thereby enhancing its overall performance. AHNPL also introduces a contrastive learning approach using a multimodal hard negative loss to improve the model's discrimination of hard negatives within each modality and a dynamic margin loss that adjusts the contrastive margin according to sample difficulty to enhance the distinction of challenging sample pairs. Experiments on three public datasets demonstrate that our method effectively boosts VLMs' performance on complex CR tasks. The source code is available at https://github.com/nynu-BDAI/AHNPL.

Zhongchun Zhou, Yuhang Gu, Chengtao Lai et al.

To efficiently support Large Language Models (LLMs), modern GPGPU architectures have introduced new features and programming paradigms, such as warp specialization. These features enable temporal overlap between the producer and consumer, as well as between matrix multiplication and activation function operations, substantially improving performance. To conduct effective AI infrastructure and computer architecture research, cycle-accurate simulators that support these new features, together with analytical models that faithfully capture workload characteristics, are essential. However, existing academic tools provide limited support for these emerging requirements. Existing cycle-accurate simulators do not incorporate new NVIDIA GPU features, such as the Tensor Memory Accelerator (TMA), in a timely manner. Moreover, existing analytical models can misestimate DRAM traffic under certain configurations. In this paper, we build a simulation pipeline from FlashAttention-3 kernel instrumentation to cycle-accurate simulation. The simulator achieves a mean absolute percentage error (MAPE) of 5.7\% and a maximum absolute percentage error of 12.7\% against H800. We also provide a theoretical analysis of FlashAttention-3 and explain why existing analytical models can produce inaccurate traffic estimates.

Hongzhi Huang, Defa Zhu, Banggu Wu et al. · bytedance

Tokenization is a fundamental component of large language models (LLMs), yet its influence on model scaling and performance is not fully explored. In this paper, we introduce Over-Tokenized Transformers, a novel framework that decouples input and output vocabularies to improve language modeling performance. Specifically, our approach scales up input vocabularies to leverage multi-gram tokens. Through extensive experiments, we uncover a log-linear relationship between input vocabulary size and training loss, demonstrating that larger input vocabularies consistently enhance model performance, regardless of model size. Using a large input vocabulary, we achieve performance comparable to double-sized baselines with no additional cost. Our findings highlight the importance of tokenization in scaling laws and provide practical insight for tokenizer design, paving the way for more efficient and powerful LLMs.

Jiacong Wang, Zijian Kang, Haochen Wang et al.

In the field of multimodal chain-of-thought (CoT) reasoning, existing approaches predominantly rely on reasoning on pure language space, which inherently suffers from language bias and is largely confined to math or science domains. This narrow focus limits their ability to handle complex visual reasoning tasks that demand comprehensive understanding of image details. To address these limitations, this paper introduces VGR, a novel reasoning multimodal large language model (MLLM) with enhanced fine-grained visual perception capabilities. Unlike traditional MLLMs that answer the question or reasoning solely on the language space, our VGR first detects relevant regions that may help to solve problems, and then provides precise answers based on replayed image regions. To achieve this, we conduct a large-scale SFT dataset called VGR -SFT that contains reasoning data with mixed vision grounding and language deduction. The inference pipeline of VGR allows the model to choose bounding boxes for visual reference and a replay stage is introduced to integrates the corresponding regions into the reasoning process, enhancing multimodel comprehension. Experiments on the LLaVA-NeXT-7B baseline show that VGR achieves superior performance on multi-modal benchmarks requiring comprehensive image detail understanding. Compared to the baseline, VGR uses only 30\% of the image token count while delivering scores of +4.1 on MMStar, +7.1 on AI2D, and a +12.9 improvement on ChartQA.

Ya Wang, Guangzheng Zhu, Cungen Cao et al.

Knowledge about emotional events is an important kind of knowledge which has been applied to improve the effectiveness of different applications. However, emotional events cannot be easily acquired, especially common or generalized emotional events that are context-independent. The goal of this paper is to obtain common emotional events in Chinese language such as "win a prize" and "be criticized". Our approach begins by collecting a comprehensive list of Chinese emotional event indicators. Then, we generate emotional events by prompting a Chinese large language model (LLM) using these indicators. To ensure the quality of these emotional events, we train a filter to discard invalid generated results. We also classify these emotional events as being positive events and negative events using different techniques. Finally, we harvest a total of 102,218 high-quality common emotional events with sentiment polarity labels, which is the only large-scale commonsense knowledge base of emotional events in Chinese language. Intrinsic evaluation results show that the proposed method in this paper can be effectively used to acquire common Chinese emotional events. An extrinsic use case also demonstrates the strong potential of common emotional events in the field of emotion cause extraction (ECE). Related resources including emotional event indicators and emotional events will be released after the publication of this paper.

Wei Zhang, Jian Yang, Jiaxi Yang et al.

Code large language models (LLMs) enhance programming by understanding and generating code across languages, offering intelligent feedback, bug detection, and code updates through reflection, improving development efficiency and accessibility. While benchmarks (e.g. HumanEval/LiveCodeBench) evaluate code generation and real-world relevance, previous works ignore the scenario of modifying code in repositories. Considering challenges remaining in improving reflection capabilities and avoiding data contamination in dynamic benchmarks, we introduce LiveRepoReflection, a challenging benchmark for evaluating code understanding and generation in multi-file repository contexts, featuring 1,888 rigorously filtered test cases across $6$ programming languages to ensure diversity, correctness, and high difficulty. Further, we create RepoReflection-Instruct, a large-scale, quality-filtered instruction-tuning dataset derived from diverse sources, used to train RepoReflectionCoder through a two-turn dialogue process involving code generation and error-driven repair. The leaderboard evaluates over 40 LLMs to reflect the model performance of repository-based code reflection.

Bohong Wu, Shen Yan, Sijun Zhang et al. · bytedance

Recent advances in large language models have demonstrated the effectiveness of length scaling during post-training, yet its potential in pre-training remains underexplored. We present the Parallel Hidden Decoding Transformer (\textit{PHD}-Transformer), a novel framework that enables efficient length scaling during pre-training while maintaining inference efficiency. \textit{PHD}-Transformer achieves this through an innovative KV cache management strategy that distinguishes between original tokens and hidden decoding tokens. By retaining only the KV cache of original tokens for long-range dependencies while immediately discarding hidden decoding tokens after use, our approach maintains the same KV cache size as the vanilla transformer while enabling effective length scaling. To further enhance performance, we introduce two optimized variants: \textit{PHD-SWA} employs sliding window attention to preserve local dependencies, while \textit{PHD-CSWA} implements chunk-wise sliding window attention to eliminate linear growth in pre-filling time. Extensive experiments demonstrate consistent improvements across multiple benchmarks.

Ziming Chen, L. Ruby Leung, Wenyu Zhou et al.

Machine learning (ML)-based models have demonstrated high skill and computational efficiency, often outperforming conventional physics-based models in weather and subseasonal predictions. While prior studies have assessed their fidelity in capturing synoptic-scale atmospheric dynamics, their performance across timescales and under out-of-distribution forcing, such as +3K or +4K uniform-warming forcings, and the sources of biases remain elusive, to establish the model reliability for Earth science. Here, we design three sets of experiments targeting synoptic-scale phenomena, interannual variability, and out-of-distribution uniform-warming forcings. We evaluate the Neural General Circulation Model (NeuralGCM), a hybrid model integrating a dynamical core with ML-based component, against observations and physics-based Earth system models (ESMs). At the synoptic scale, NeuralGCM captures the evolution and propagation of extratropical cyclones with performance comparable to ESMs. At the interannual scale, when forced by El Niño-Southern Oscillation sea surface temperature (SST) anomalies, NeuralGCM successfully reproduces associated teleconnection patterns but exhibits deficiencies in capturing nonlinear response. Under out-of-distribution uniform-warming forcings, NeuralGCM simulates similar responses in global-average temperature and precipitation and reproduces large-scale tropospheric circulation features similar to those in ESMs. Notable weaknesses include overestimating the tracks and spatial extent of extratropical cyclones, biases in the teleconnected wave train triggered by tropical SST anomalies, and differences in upper-level warming and stratospheric circulation responses to SST warming compared to physics-based ESMs. The causes of these weaknesses were explored.

Ya Wang, Hesen Chen, Fangyi Zhang et al.

Data augmentation is a commonly used approach to improving the generalization of deep learning models. Recent works show that learned data augmentation policies can achieve better generalization than hand-crafted ones. However, most of these works use unified augmentation policies for all samples in a dataset, which is observed not necessarily beneficial for all labels in multi-label classification tasks, i.e., some policies may have negative impacts on some labels while benefitting the others. To tackle this problem, we propose a novel Label-Based AutoAugmentation (LB-Aug) method for multi-label scenarios, where augmentation policies are generated with respect to labels by an augmentation-policy network. The policies are learned via reinforcement learning using policy gradient methods, providing a mapping from instance labels to their optimal augmentation policies. Numerical experiments show that our LB-Aug outperforms previous state-of-the-art augmentation methods by large margins in multiple benchmarks on image and video classification.

Zhenzhen Gu, Cungen Cao, Ya Wang et al.

To make machine exhibit human-like abilities in the domains like robotics and conversation, social commonsense knowledge (SCK), i.e., common sense about social contexts and social roles, is absolutely necessarily. Therefor, our ultimate goal is to acquire large-scale SCK to support much more intelligent applications. Before that, we need to know clearly what is SCK and how to represent it, since automatic information processing requires data and knowledge are organized in structured and semantically related ways. For this reason, in this paper, we identify and formalize three basic types of SCK based on first-order theory. Firstly, we identify and formalize the interrelationships, such as having-role and having-social_relation, among social contexts, roles and players from the perspective of considering both contexts and roles as first-order citizens and not generating role instances. Secondly, we provide a four level structure to identify and formalize the intrinsic information, such as events and desires, of social contexts, roles and players, and illustrate the way of harvesting the intrinsic information of social contexts and roles from the exhibition of players in concrete contexts. And thirdly, enlightened by some observations of actual contexts, we further introduce and formalize the embedding of social contexts, and depict the way of excavating the intrinsic information of social contexts and roles from the embedded smaller and simpler contexts. The results of this paper lay the foundation not only for formalizing much more complex SCK but also for acquiring these three basic types of SCK.

Ya Wang, Dongliang He, Fu Li et al.

Images or videos always contain multiple objects or actions. Multi-label recognition has been witnessed to achieve pretty performance attribute to the rapid development of deep learning technologies. Recently, graph convolution network (GCN) is leveraged to boost the performance of multi-label recognition. However, what is the best way for label correlation modeling and how feature learning can be improved with label system awareness are still unclear. In this paper, we propose a label graph superimposing framework to improve the conventional GCN+CNN framework developed for multi-label recognition in the following two aspects. Firstly, we model the label correlations by superimposing label graph built from statistical co-occurrence information into the graph constructed from knowledge priors of labels, and then multi-layer graph convolutions are applied on the final superimposed graph for label embedding abstraction. Secondly, we propose to leverage embedding of the whole label system for better representation learning. In detail, lateral connections between GCN and CNN are added at shallow, middle and deep layers to inject information of label system into backbone CNN for label-awareness in the feature learning process. Extensive experiments are carried out on MS-COCO and Charades datasets, showing that our proposed solution can greatly improve the recognition performance and achieves new state-of-the-art recognition performance.