Xiaotong Li

Meituan LongCat Team, Bin Xiao, Chao Wang et al.

The prevailing Next-Token Prediction (NTP) paradigm has driven the success of large language models through discrete autoregressive modeling. However, contemporary multimodal systems remain language-centric, often treating non-linguistic modalities as external attachments, leading to fragmented architectures and suboptimal integration. To transcend this limitation, we introduce Discrete Native Autoregressive (DiNA), a unified framework that represents multimodal information within a shared discrete space, enabling a consistent and principled autoregressive modeling across modalities. A key innovation is the Discrete Native Any-resolution Visual Transformer (dNaViT), which performs tokenization and de-tokenization at arbitrary resolutions, transforming continuous visual signals into hierarchical discrete tokens. Building on this foundation, we develop LongCat-Next, a native multimodal model that processes text, vision, and audio under a single autoregressive objective with minimal modality-specific design. As an industrial-strength foundation model, it excels at seeing, painting, and talking within a single framework, achieving strong performance across a wide range of multimodal benchmarks. In particular, LongCat-Next addresses the long-standing performance ceiling of discrete vision modeling on understanding tasks and provides a unified approach to effectively reconcile the conflict between understanding and generation. As an attempt toward native multimodality, we open-source the LongCat-Next and its tokenizers, hoping to foster further research and development in the community. GitHub: https://github.com/meituan-longcat/LongCat-Next

Haiwen Diao, Xiaotong Li, Yufeng Cui et al.

Existing encoder-free vision-language models (VLMs) are rapidly narrowing the performance gap with their encoder-based counterparts, highlighting the promising potential for unified multimodal systems with structural simplicity and efficient deployment. We systematically clarify the performance gap between VLMs using pre-trained vision encoders, discrete tokenizers, and minimalist visual layers from scratch, deeply excavating the under-examined characteristics of encoder-free VLMs. We develop efficient strategies for encoder-free VLMs that rival mainstream encoder-based ones. After an in-depth investigation, we launch EVEv2.0, a new and improved family of encoder-free VLMs. We show that: (i) Properly decomposing and hierarchically associating vision and language within a unified model reduces interference between modalities. (ii) A well-designed training strategy enables effective optimization for encoder-free VLMs. Through extensive evaluation, our EVEv2.0 represents a thorough study for developing a decoder-only architecture across modalities, demonstrating superior data efficiency and strong vision-reasoning capability. Code is publicly available at: https://github.com/baaivision/EVE.

Xiaotong Li, Gang Chen, Javier Alonso-Mora

Open-world object manipulation remains a fundamental challenge in robotics. While Vision-Language-Action (VLA) models have demonstrated promising results, they rely heavily on large-scale robot action demonstrations, which are costly to collect and can hinder out-of-distribution generalization. In this paper, we propose an explicit-world-model-based framework for open-world manipulation that achieves zero-shot generalization by constructing a physically grounded digital twin of the environment. The framework integrates open-set perception, digital-twin reconstruction, sampling and evaluation of interaction strategies. By constructing a digital twin of the environment, our approach efficiently explores and evaluates manipulation strategies in physic-enabled simulator and reliably deploys the chosen strategy to the real world. Experimentally, the proposed framework is able to perform multiple open-set manipulation tasks without any task-specific action demonstrations, proving strong zero-shot generalization on both the task and object levels. Project Page: https://bojack-bj.github.io/projects/thesis/

Haiwen Diao, Yufeng Cui, Xiaotong Li et al.

Existing vision-language models (VLMs) mostly rely on vision encoders to extract visual features followed by large language models (LLMs) for visual-language tasks. However, the vision encoders set a strong inductive bias in abstracting visual representation, e.g., resolution, aspect ratio, and semantic priors, which could impede the flexibility and efficiency of the VLMs. Training pure VLMs that accept the seamless vision and language inputs, i.e., without vision encoders, remains challenging and rarely explored. Empirical observations reveal that direct training without encoders results in slow convergence and large performance gaps. In this work, we bridge the gap between encoder-based and encoder-free models, and present a simple yet effective training recipe towards pure VLMs. Specifically, we unveil the key aspects of training encoder-free VLMs efficiently via thorough experiments: (1) Bridging vision-language representation inside one unified decoder; (2) Enhancing visual recognition capability via extra supervision. With these strategies, we launch EVE, an encoder-free vision-language model that can be trained and forwarded efficiently. Notably, solely utilizing 35M publicly accessible data, EVE can impressively rival the encoder-based VLMs of similar capacities across multiple vision-language benchmarks. It significantly outperforms the counterpart Fuyu-8B with mysterious training procedures and undisclosed training data. We believe that EVE provides a transparent and efficient route for developing a pure decoder-only architecture across modalities. Our code and models are publicly available at: https://github.com/baaivision/EVE.

Zixuan Hu, Xiaotong Li, Shixiang Tang et al.

The remarkable success of pretrain-then-finetune paradigm has led to a proliferation of available pre-trained models for vision tasks. This surge presents a significant challenge in efficiently choosing the most suitable pre-trained models for downstream tasks. The critical aspect of this challenge lies in effectively predicting the model transferability by considering the underlying fine-tuning dynamics. Existing methods often model fine-tuning dynamics in feature space with linear transformations, which do not precisely align with the fine-tuning objective and fail to grasp the essential nonlinearity from optimization. To this end, we present LEAD, a finetuning-aligned approach based on the network output of logits. LEAD proposes a theoretical framework to model the optimization process and derives an ordinary differential equation (ODE) to depict the nonlinear evolution toward the final logit state. Additionally, we design a class-aware decomposition method to consider the varying evolution dynamics across classes and further ensure practical applicability. Integrating the closely aligned optimization objective and nonlinear modeling capabilities derived from the differential equation, our method offers a concise solution to effectively bridge the optimization gap in a single step, bypassing the lengthy fine-tuning process. The comprehensive experiments on 24 supervised and self-supervised pre-trained models across 10 downstream datasets demonstrate impressive performances and showcase its broad adaptability even in low-data scenarios.

Yongxing Dai, Xiaotong Li, Jun Liu et al.

Domain generalizable (DG) person re-identification (ReID) is a challenging problem because we cannot access any unseen target domain data during training. Almost all the existing DG ReID methods follow the same pipeline where they use a hybrid dataset from multiple source domains for training, and then directly apply the trained model to the unseen target domains for testing. These methods often neglect individual source domains' discriminative characteristics and their relevances w.r.t. the unseen target domains, though both of which can be leveraged to help the model's generalization. To handle the above two issues, we propose a novel method called the relevance-aware mixture of experts (RaMoE), using an effective voting-based mixture mechanism to dynamically leverage source domains' diverse characteristics to improve the model's generalization. Specifically, we propose a decorrelation loss to make the source domain networks (experts) keep the diversity and discriminability of individual domains' characteristics. Besides, we design a voting network to adaptively integrate all the experts' features into the more generalizable aggregated features with domain relevance. Considering the target domains' invisibility during training, we propose a novel learning-to-learn algorithm combined with our relation alignment loss to update the voting network. Extensive experiments demonstrate that our proposed RaMoE outperforms the state-of-the-art methods.