Yulong Ao

Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo et al. · tsinghua

While next-token prediction is considered a promising path towards artificial general intelligence, it has struggled to excel in multimodal tasks, which are still dominated by diffusion models (e.g., Stable Diffusion) and compositional approaches (e.g., CLIP combined with LLMs). In this paper, we introduce Emu3, a new suite of state-of-the-art multimodal models trained solely with next-token prediction. By tokenizing images, text, and videos into a discrete space, we train a single transformer from scratch on a mixture of multimodal sequences. Emu3 outperforms several well-established task-specific models in both generation and perception tasks, surpassing flagship models such as SDXL and LLaVA-1.6, while eliminating the need for diffusion or compositional architectures. Emu3 is also capable of generating high-fidelity video via predicting the next token in a video sequence. We simplify complex multimodal model designs by converging on a singular focus: tokens, unlocking great potential for scaling both during training and inference. Our results demonstrate that next-token prediction is a promising path towards building general multimodal intelligence beyond language. We open-source key techniques and models to support further research in this direction.

Yuejie Wang, Tao Chang, Yuanyuan Zhao et al.

Training Large Language Models (LLMs) on heterogeneous clusters presents significant challenges for collective communication, as hardware from multiple vendors introduces diverse network and computational characteristics. Existing collective communication frameworks (e.g., NCCL, RCCL) designed for homogeneous environments fail to address mixed-hardware setups, while communication libraries with heterogeneous support (e.g., Gloo, OpenMPI) incur heavy overhead in the data path. This paper presents HetCCL, a framework that enables heterogeneous collective communication by efficient P2P transport across heterogeneous devices (e.g., GPUs), eliminating the host-device memory copy overhead while offloading the control to the CPUs. For combining collectives (e.g., AllReduce, ReduceScatter), HetCCL introduces a border-communicator mechanism that achieves vendor independence by using the intrinsic reduction in the combining collectives in vendor collective communication libraries. With efficient heterogeneous P2P transport and portable reduction mechanism, HetCCL proposes a hierarchical topology abstraction for heterogeneous clusters, dissecting collective communication into cluster-level primitives that guarantee optimal cross-cluster data transfer volume and optimal bandwidth utilization. We implement HetCCL with 4 different vendor support and evaluate it in 4 heterogeneous settings with benchmarks and end-to-end LLM tasks. Our evaluation shows that HetCCL achieves 17-19x higher bandwidth than Gloo in heterogeneous communications, and speeds up end-to-end training by up to 16.9% in the per-step-time.

Lu Wang, Zhuoran Jin, Yupu Hao et al.

Multimodal large language models (MLLMs) have shown strong performance on offline video understanding, but most are limited to offline inference or have weak online reasoning, making multi-turn interaction over continuously arriving video streams difficult. Existing streaming methods typically use an interleaved perception-generation paradigm, which prevents concurrent perception and generation and leads to early memory decay as streams grow, hurting long-range dependency modeling. We propose Think While Watching, a memory-anchored streaming video reasoning framework that preserves continuous segment-level memory during multi-turn interaction. We build a three-stage, multi-round chain-of-thought dataset and adopt a stage-matched training strategy, while enforcing strict causality through a segment-level streaming causal mask and streaming positional encoding. During inference, we introduce an efficient pipeline that overlaps watching and thinking and adaptively selects the best attention backend. Under both single-round and multi-round streaming input protocols, our method achieves strong results. Built on Qwen3-VL, it improves single-round accuracy by 2.6% on StreamingBench and by 3.79% on OVO-Bench. In the multi-round setting, it maintains performance while reducing output tokens by 56%. Code is available at: https://github.com/wl666hhh/Think_While_Watching/

Yufeng Cui, Honghao Chen, Haoge Deng et al.

We introduce Emu3.5, a large-scale multimodal world model that natively predicts the next state across vision and language. Emu3.5 is pre-trained end-to-end with a unified next-token prediction objective on a corpus of vision-language interleaved data containing over 10 trillion tokens, primarily derived from sequential frames and transcripts of internet videos. The model naturally accepts interleaved vision-language inputs and generates interleaved vision-language outputs. Emu3.5 is further post-trained with large-scale reinforcement learning to enhance multimodal reasoning and generation. To improve inference efficiency, we propose Discrete Diffusion Adaptation (DiDA), which converts token-by-token decoding into bidirectional parallel prediction, accelerating per-image inference by about 20x without sacrificing performance. Emu3.5 exhibits strong native multimodal capabilities, including long-horizon vision-language generation, any-to-image (X2I) generation, and complex text-rich image generation. It also exhibits generalizable world-modeling abilities, enabling spatiotemporally consistent world exploration and open-world embodied manipulation across diverse scenarios and tasks. For comparison, Emu3.5 achieves performance comparable to Gemini 2.5 Flash Image (Nano Banana) on image generation and editing tasks and demonstrates superior results on a suite of interleaved generation tasks. We open-source Emu3.5 at https://github.com/baaivision/Emu3.5 to support community research.

Bo-Wen Zhang, Liangdong Wang, Jijie Li et al.

This paper introduces the Aquila2 series, which comprises a wide range of bilingual models with parameter sizes of 7, 34, and 70 billion. These models are trained based on an innovative framework named HeuriMentor (HM), which offers real-time insights into model convergence and enhances the training process and data management. The HM System, comprising the Adaptive Training Engine (ATE), Training State Monitor (TSM), and Data Management Unit (DMU), allows for precise monitoring of the model's training progress and enables efficient optimization of data distribution, thereby enhancing training effectiveness. Extensive evaluations show that the Aquila2 model series performs comparably well on both English and Chinese benchmarks. Specifically, Aquila2-34B demonstrates only a slight decrease in performance when quantized to Int4. Furthermore, we have made our training code (https://github.com/FlagOpen/FlagScale) and model weights (https://github.com/FlagAI-Open/Aquila2) publicly available to support ongoing research and the development of applications.

Shuhao Gu, Jialing Zhang, Siyuan Zhou et al.

Recently, Vision-Language Models (VLMs) have achieved remarkable progress in multimodal tasks, and multimodal instruction data serves as the foundation for enhancing VLM capabilities. Despite the availability of several open-source multimodal datasets, limitations in the scale and quality of open-source instruction data hinder the performance of VLMs trained on these datasets, leading to a significant gap compared to models trained on closed-source data. To address this challenge, we introduce Infinity-MM, a large-scale multimodal instruction dataset. We collected the available multimodal instruction datasets and performed unified preprocessing, resulting in a dataset with over 40 million samples that ensures diversity and accuracy. Furthermore, to enable large-scale expansion of instruction data and support the continuous acquisition of high-quality data, we propose a synthetic instruction generation method based on a tagging system and open-source VLMs. By establishing correspondences between different types of images and associated instruction types, this method can provide essential guidance during data synthesis. Leveraging this high-quality data, we have trained a 2-billion-parameter Vision-Language Model, Aquila-VL-2B, which achieves state-of-the-art (SOTA) performance among models of similar scale. The data is available at: https://huggingface.co/datasets/BAAI/Infinity-MM.

Bo-Wen Zhang, Liangdong Wang, Ye Yuan et al.

In recent years, with the rapid application of large language models across various fields, the scale of these models has gradually increased, and the resources required for their pre-training have grown exponentially. Training an LLM from scratch will cost a lot of computation resources while scaling up from a smaller model is a more efficient approach and has thus attracted significant attention. In this paper, we present AquilaMoE, a cutting-edge bilingual 8*16B Mixture of Experts (MoE) language model that has 8 experts with 16 billion parameters each and is developed using an innovative training methodology called EfficientScale. This approach optimizes performance while minimizing data requirements through a two-stage process. The first stage, termed Scale-Up, initializes the larger model with weights from a pre-trained smaller model, enabling substantial knowledge transfer and continuous pretraining with significantly less data. The second stage, Scale-Out, uses a pre-trained dense model to initialize the MoE experts, further enhancing knowledge transfer and performance. Extensive validation experiments on 1.8B and 7B models compared various initialization schemes, achieving models that maintain and reduce loss during continuous pretraining. Utilizing the optimal scheme, we successfully trained a 16B model and subsequently the 8*16B AquilaMoE model, demonstrating significant improvements in performance and training efficiency.

Hao Zhang, Mengsi Lyu, Bo Huang et al.

Large Multimodal Models (LMMs) have proven effective on various tasks. They typically encode visual inputs into Original Model sequences of tokens, which are then concatenated with textual tokens and jointly processed by the language model. However, the growing number of visual tokens greatly increases inference cost. Visual token pruning has emerged as a promising solution. However, existing methods often overlook scenarios involving long context inputs with multiple images. In this paper, we analyze the challenges of visual token pruning in long context, multi-image settings and introduce an adaptive pruning method tailored for such scenarios. We decompose redundancy into intra-image and inter-image components and quantify them through intra-image diversity and inter-image variation, which jointly guide dynamic budget allocation. Our approach consists of two stages. The intra-image stage allocates each image a content-aware token budget and greedily selects its most representative tokens. The inter-image stage performs global diversity filtering to form a candidate pool and then applies a Pareto selection procedure that balances diversity with text alignment. Extensive experiments show that our approach can reduce up to 80% of visual tokens while maintaining performance in long context settings.

Zihao Zheng, Xingyue Zhou, Zhihao Mao et al.

Vision-Language-Navigation (VLN) models exhibit excellent navigation accuracy but incur high computational overhead. Token caching has emerged as a promising training-free strategy to reduce this cost by reusing token computation results; however, existing token caching approaches rely on visual domain methods for cacheable token selection, leading to challenges when adapted to VLN models. 1) Visual domain methods become invalid when there is viewpoint migration. 2) Visual domain methods neglect critical edge information without the aid of additional algorithms. 3) Visual domain methods overlook the temporal variation of scenarios and lack adjustability in cache budgets. In this paper, we develop detailed analyses and find that the impacts of these challenges exhibit invariance and analyzability in the frequency domain. Based on these, we propose a frequency-guided token caching framework, called FreqCache. Utilizing the inherent properties of the frequency domain, FreqCache achieves optimal token cache establishment, refreshment, and adaptive adjustment. Experiments show that FreqCache achieves 1.59x speedup with ignorable overhead, showing the effect of integrating frequency domain methods in VLN token caching.

Hao Zhang, Mengsi Lyu, Chenrui He et al.

Large Multimodal Models (LMMs) have achieved significant success across various tasks. These models usually encode visual inputs into dense token sequences, which are then concatenated with textual tokens and jointly processed by a language model. However, the increased token count substantially raises computational and memory costs during inference. Token pruning has emerged as a promising approach to address this issue. Existing token pruning methods often rely on costly calibration or suboptimal importance metrics, leading to redundant retained tokens. In this paper, we analyze the redundancy differences between visual and textual tokens and propose pruning exclusively on visual tokens. Based on this, we propose a visual token pruning strategy that explicitly preserves both cross-modal alignment and intra-modal informational diversity. We introduce a mutual information-based token pruning strategy that removes visual tokens semantically misaligned with textual tokens, effectively preserving the alignment between the visual and textual modalities. To further improve the representational quality of the retained tokens, we additionally prune redundant visual tokens by maximizing the expected pairwise distances in the embedding space, which is solved efficiently with a greedy algorithm. Extensive experiments demonstrate that our method maintains strong performance while reducing tokens by 88.9% on models such as LLaVA-1.5-7B and LLaVA-NEXT-7B, resulting in a 56.7% improvement in inference speed.

Hao Zhang, Mengsi Lyu, Zhuo Chen et al.

Large Language Models (LLMs) demonstrate exceptional capabilities across various tasks, but their deployment is constrained by high computational and memory costs. Model pruning provides an effective means to alleviate these demands. However, existing methods often ignore the characteristics of prefill-decode (PD) disaggregation in practice. In this paper, we propose a novel pruning method for PD disaggregation inference, enabling more precise and efficient block and KV Cache pruning. Our approach constructs pruning and distillation sets to perform iterative block removal independently for the prefill and decode stages, obtaining better pruning solutions. Moreover, we introduce a token-aware cache pruning mechanism that retains all KV Cache in the prefill stage but selectively reuses entries for the first and last token sequences in selected layers during decode, reducing communication costs with minimal overhead. Extensive experiments demonstrate that our approach consistently achieves strong performance in both PD disaggregation and PD unified settings without disaggregation. Under the same (default) settings, our method achieves improved performance and faster inference, along with a 4.95$\times$ reduction in data transmission bandwidth consumption.

Yulong Ao, Zhihua Wu, Dianhai Yu et al.

Distributed training has become a pervasive and effective approach for training a large neural network (NN) model with processing massive data. However, it is very challenging to satisfy requirements from various NN models, diverse computing resources, and their dynamic changes during a training job. In this study, we design our distributed training framework in a systematic end-to-end view to provide the built-in adaptive ability for different scenarios, especially for industrial applications and production environments, by fully considering resource allocation, model partition, task placement, and distributed execution. Based on the unified distributed graph and the unified cluster object, our adaptive framework is equipped with a global cost model and a global planner, which can enable arbitrary parallelism, resource-aware placement, multi-mode execution, fault-tolerant, and elastic distributed training. The experiments demonstrate that our framework can satisfy various requirements from the diversity of applications and the heterogeneity of resources with highly competitive performance. The ERNIE language model with 260 billion parameters is efficiently trained on thousands of AI processors with 91.7% weak scalability. The throughput of the model from the recommender system by employing the heterogeneous pipeline asynchronous execution can be increased up to 2.1 times and 3.3 times that of the GPU-only and CPU-only training respectively. Moreover, the fault-tolerant and elastic distributed training have been successfully applied to the online industrial applications, which give a reduction of 34.49% in the number of failed long-term training jobs and an increase of 33.91% for the global scheduling efficiency in the production environment.