Jiarui Zhao

Meituan LongCat Team, Anchun Gui, Bei Li et al.

We introduce LongCat-Flash-Thinking-2601, a 560-billion-parameter open-source Mixture-of-Experts (MoE) reasoning model with superior agentic reasoning capability. LongCat-Flash-Thinking-2601 achieves state-of-the-art performance among open-source models on a wide range of agentic benchmarks, including agentic search, agentic tool use, and tool-integrated reasoning. Beyond benchmark performance, the model demonstrates strong generalization to complex tool interactions and robust behavior under noisy real-world environments. Its advanced capability stems from a unified training framework that combines domain-parallel expert training with subsequent fusion, together with an end-to-end co-design of data construction, environments, algorithms, and infrastructure spanning from pre-training to post-training. In particular, the model's strong generalization capability in complex tool-use are driven by our in-depth exploration of environment scaling and principled task construction. To optimize long-tailed, skewed generation and multi-turn agentic interactions, and to enable stable training across over 10,000 environments spanning more than 20 domains, we systematically extend our asynchronous reinforcement learning framework, DORA, for stable and efficient large-scale multi-environment training. Furthermore, recognizing that real-world tasks are inherently noisy, we conduct a systematic analysis and decomposition of real-world noise patterns, and design targeted training procedures to explicitly incorporate such imperfections into the training process, resulting in improved robustness for real-world applications. To further enhance performance on complex reasoning tasks, we introduce a Heavy Thinking mode that enables effective test-time scaling by jointly expanding reasoning depth and width through intensive parallel thinking.

Yukuan Zhang, Jiarui Zhao, Shangqing Nie et al.

Cross-modal hashing retrieval encodes heterogeneous data into compact binary codes for efficient Hamming-space search. Existing methods usually learn cross-modal semantics in continuous feature spaces and generate binary codes through a final sign operation, which weakly couples training optimization with discrete hash retrieval. We propose SpikeHash, a unified spiking framework that formulates cross-modal hashing as spike-state evolution, directional spike interaction, and competitive spike readout. Specifically, SpikeHash converts image and text features into multi-timestep spike sequences. In a shared Hamming space, the two spike sequences jointly drive the temporal evolution of a shared hash state. Cross-modal interaction is further performed through directional spike modulation, enabling each modality to influence the firing dynamics of the other. Crucially, SpikeHash replaces the conventional continuous hash head with a positive-negative spiking hash readout, where each hash bit is produced by temporal competition between paired spike channels. Experimental results show that SpikeHash achieves competitive retrieval accuracy on three benchmark datasets while reducing the parameter size, operation count, and estimated energy of the hash learning stage, suggesting a compact spiking alternative to conventional continuous hash mapping. The project page is available at https://shuqiao-111.github.io/.

Yu Wang, Chuanguang Yang, Zhulin An et al.

Existing Knowledge Distillation (KD) methods often align feature information between teacher and student by exploring meaningful feature processing and loss functions. However, due to the difference in feature distributions between the teacher and student, the student model may learn incompatible information from the teacher. To address this problem, we propose teacher-guided student Diffusion Self-KD, dubbed as DSKD. Instead of the direct teacher-student alignment, we leverage the teacher classifier to guide the sampling process of denoising student features through a light-weight diffusion model. We then propose a novel locality-sensitive hashing (LSH)-guided feature distillation method between the original and denoised student features. The denoised student features encapsulate teacher knowledge and could be regarded as a teacher role. In this way, our DSKD method could eliminate discrepancies in mapping manners and feature distributions between the teacher and student, while learning meaningful knowledge from the teacher. Experiments on visual recognition tasks demonstrate that DSKD significantly outperforms existing KD methods across various models and datasets. Our code is attached in supplementary material.

Meihua Zhou, Liping Yu, Jiawei Cai et al.

Hyperspectral image (HSI) classification typically involves large-scale data and computationally intensive training, which limits the practical deployment of deep learning models in real-world remote sensing tasks. This study introduces SpectralTrain, a universal, architecture-agnostic training framework that enhances learning efficiency by integrating curriculum learning (CL) with principal component analysis (PCA)-based spectral downsampling. By gradually introducing spectral complexity while preserving essential information, SpectralTrain enables efficient learning of spectral -- spatial patterns at significantly reduced computational costs. The framework is independent of specific architectures, optimizers, or loss functions and is compatible with both classical and state-of-the-art (SOTA) models. Extensive experiments on three benchmark datasets -- Indian Pines, Salinas-A, and the newly introduced CloudPatch-7 -- demonstrate strong generalization across spatial scales, spectral characteristics, and application domains. The results indicate consistent reductions in training time by 2-7x speedups with small-to-moderate accuracy deltas depending on backbone. Its application to cloud classification further reveals potential in climate-related remote sensing, emphasizing training strategy optimization as an effective complement to architectural design in HSI models. Code is available at https://github.com/mh-zhou/SpectralTrain.

Meihua Zhou, Xinyu Tong, Jiarui Zhao et al.

High-dimensional neuroimaging analyses for clinical diagnosis are often constrained by compromises in spatiotemporal fidelity and by the limited adaptability of large-scale, general-purpose models. To address these challenges, we introduce Dynamic Curriculum Learning for Spatiotemporal Encoding (DCL-SE), an end-to-end framework centered on data-driven spatiotemporal encoding (DaSE). We leverage Approximate Rank Pooling (ARP) to efficiently encode three-dimensional volumetric brain data into information-rich, two-dimensional dynamic representations, and then employ a dynamic curriculum learning strategy, guided by a Dynamic Group Mechanism (DGM), to progressively train the decoder, refining feature extraction from global anatomical structures to fine pathological details. Evaluated across six publicly available datasets, including Alzheimer's disease and brain tumor classification, cerebral artery segmentation, and brain age prediction, DCL-SE consistently outperforms existing methods in accuracy, robustness, and interpretability. These findings underscore the critical importance of compact, task-specific architectures in the era of large-scale pretrained networks.

Yukuan Zhang, Jiarui Zhao, Shangqing Nie et al.

Multimodal semantic cues, such as textual descriptions, have shown strong potential in enhancing target perception for tracking. However, existing methods rely on static textual descriptions from large language models, which lack adaptability to real-time target state changes and prone to hallucinations. To address these challenges, we propose a unified multimodal vision-language tracking framework, named EPIPTrack, which leverages explicit and implicit prompts for dynamic target modeling and semantic alignment. Specifically, explicit prompts transform spatial motion information into natural language descriptions to provide spatiotemporal guidance. Implicit prompts combine pseudo-words with learnable descriptors to construct individualized knowledge representations capturing appearance attributes. Both prompts undergo dynamic adjustment via the CLIP text encoder to respond to changes in target state. Furthermore, we design a Discriminative Feature Augmentor to enhance visual and cross-modal representations. Extensive experiments on MOT17, MOT20, and DanceTrack demonstrate that EPIPTrack outperforms existing trackers in diverse scenarios, exhibiting robust adaptability and superior performance.