Yajie Zhang

Yuanyang Yin, Yaqi Zhao, Yajie Zhang et al.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities by integrating visual and textual inputs, yet modality alignment remains one of the most challenging aspects. Current MLLMs typically rely on simple adapter architectures and pretraining approaches to bridge vision encoders with large language models (LLM), guided by image-level supervision. We identify this paradigm often leads to suboptimal alignment between modalities, significantly constraining the LLM's ability to properly interpret and reason with visual features particularly for smaller language models. This limitation degrades overall performance-particularly for smaller language models where capacity constraints are more pronounced and adaptation capabilities are limited. To address this fundamental limitation, we propose Supervised Embedding Alignment (SEA), a token-level supervision alignment method that enables more precise visual-text alignment during pretraining. SEA introduces minimal computational overhead while preserving language capabilities and substantially improving cross-modal understanding. Our comprehensive analyses reveal critical insights into the adapter's role in multimodal integration, and extensive experiments demonstrate that SEA consistently improves performance across various model sizes, with smaller models benefiting the most (average performance gain of 7.61% for Gemma-2B). This work establishes a foundation for developing more effective alignment strategies for future multimodal systems.

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

We present LongCat-Flash-Thinking, an efficient 560-billion-parameter open-source Mixture-of-Experts (MoE) reasoning model. Its advanced capabilities are cultivated through a meticulously crafted training process, beginning with long Chain-of-Thought (CoT) data cold-start and culminating in large-scale Reinforcement Learning (RL). We first employ a well-designed cold-start training strategy, which significantly enhances the reasoning potential and equips the model with specialized skills in both formal and agentic reasoning. Then, a core innovation is our domain-parallel training scheme, which decouples optimization across distinct domains (e.g., STEM, Code, Agentic) and subsequently fuses the resulting expert models into a single, nearly Pareto-optimal model. This entire process is powered by our Dynamic ORchestration for Asynchronous rollout (DORA) system, a large-scale RL framework that delivers a greater than threefold training speedup over synchronous methods on tens of thousands of accelerators. As a result, LongCat-Flash-Thinking achieves state-of-the-art performance among open-source models on a suite of complex reasoning tasks. The model exhibits exceptional efficiency in agentic reasoning, reducing average token consumption by 64.5% (from 19, 653 to 6, 965) on AIME-25, without degrading task accuracy. We release LongCat-Flash-Thinking to promote further advances in reasoning systems and agentic AI research.

Yinuo Liu, Zi Qian, Heng Zhou et al.

Interleaved text-and-image generation represents a significant frontier for Multimodal Large Language Models (MLLMs), offering a more intuitive way to convey complex information. Current paradigms rely on either image generation or retrieval augmentation, yet they typically treat the two as mutually exclusive paths, failing to unify factuality with creativity. We argue that the next milestone in this field is Agentic Tool Planning, where the model serves as a central controller that autonomously determines when, where, and which tools to invoke to produce interleaved responses for visual-critical queries. To systematically evaluate this paradigm, we introduce ATP-Bench, a novel benchmark comprising 7,702 QA pairs (including 1,592 VQA pairs) across eight categories and 25 visual-critical intents, featuring human-verified queries and ground truths. Furthermore, to evaluate agentic planning independent of end-to-end execution and changing tool backends, we propose a Multi-Agent MLLM-as-a-Judge (MAM) system. MAM evaluates tool-call precision, identifies missed opportunities for tool use, and assesses overall response quality without requiring ground-truth references. Our extensive experiments on 10 state-of-the-art MLLMs reveal that models struggle with coherent interleaved planning and exhibit significant variations in tool-use behavior, highlighting substantial room for improvement and providing actionable guidance for advancing interleaved generation. Dataset and code are available at https://github.com/Qwen-Applications/ATP-Bench.

Chonghan Liu, Haoran Wang, Felix Henry et al.

Spatial perception and reasoning are core components of human cognition, encompassing object recognition, spatial relational understanding, and dynamic reasoning. Despite progress in computer vision, existing benchmarks reveal significant gaps in models' abilities to accurately recognize object attributes and reason about spatial relationships, both essential for dynamic reasoning. To address these limitations, we propose MIRAGE, a multi-modal benchmark designed to evaluate models' capabilities in Counting (object attribute recognition), Relation (spatial relational reasoning), and Counting with Relation. Through diverse and complex scenarios requiring fine-grained recognition and reasoning, MIRAGE highlights critical limitations in state-of-the-art models, underscoring the need for improved representations and reasoning frameworks. By targeting these foundational abilities, MIRAGE provides a pathway toward spatiotemporal reasoning in future research.

Yajie Zhang, Ce Yu, Chao Sun et al.

In the astronomical observation field, determining the allocation of observation resources of the telescope array and planning follow-up observations for targets of opportunity (ToOs) are indispensable components of astronomical scientific discovery. This problem is computationally challenging, given the online observation setting and the abundance of time-varying factors that can affect whether an observation can be conducted. This paper presents ROARS, a reinforcement learning approach for online astronomical resource-constrained scheduling. To capture the structure of the astronomical observation scheduling, we depict every schedule using a directed acyclic graph (DAG), illustrating the dependency of timing between different observation tasks within the schedule. Deep reinforcement learning is used to learn a policy that can improve the feasible solution by iteratively local rewriting until convergence. It can solve the challenge of obtaining a complete solution directly from scratch in astronomical observation scenarios, due to the high computational complexity resulting from numerous spatial and temporal constraints. A simulation environment is developed based on real-world scenarios for experiments, to evaluate the effectiveness of our proposed scheduling approach. The experimental results show that ROARS surpasses 5 popular heuristics, adapts to various observation scenarios and learns effective strategies with hindsight.

Xin Xia, Yajie Zhang, Xiangxiang Zeng et al.

Molecular optimization, which aims to discover improved molecules from a vast chemical search space, is a critical step in chemical development. Various artificial intelligence technologies have demonstrated high effectiveness and efficiency on molecular optimization tasks. However, few of these technologies focus on balancing property optimization with constraint satisfaction, making it difficult to obtain high-quality molecules that not only possess desirable properties but also meet various constraints. To address this issue, we propose a constrained multi-property molecular optimization framework (CMOMO), which is a flexible and efficient method to simultaneously optimize multiple molecular properties while satisfying several drug-like constraints. CMOMO improves multiple properties of molecules with constraints based on dynamic cooperative optimization, which dynamically handles the constraints across various scenarios. Besides, CMOMO evaluates multiple properties within discrete chemical spaces cooperatively with the evolution of molecules within an implicit molecular space to guide the evolutionary search. Experimental results show the superior performance of the proposed CMOMO over five state-of-the-art molecular optimization methods on two benchmark tasks of simultaneously optimizing multiple non-biological activity properties while satisfying two structural constraints. Furthermore, the practical applicability of CMOMO is verified on two practical tasks, where it identified a collection of candidate ligands of $β$2-adrenoceptor GPCR and candidate inhibitors of glycogen synthase kinase-3$β$ with high properties and under drug-like constraints.

Honghe Zhu, Cong Wang, Yajie Zhang et al.

Single image deraining is an urgent task because the degraded rainy image makes many computer vision systems fail to work, such as video surveillance and autonomous driving. So, deraining becomes important and an effective deraining algorithm is needed. In this paper, we propose a novel network based on physical model guided learning for single image deraining, which consists of three sub-networks: rain streaks network, rain-free network, and guide-learning network. The concatenation of rain streaks and rain-free image that are estimated by rain streaks network, rain-free network, respectively, is input to the guide-learning network to guide further learning and the direct sum of the two estimated images is constrained with the input rainy image based on the physical model of rainy image. Moreover, we further develop the Multi-Scale Residual Block (MSRB) to better utilize multi-scale information and it is proved to boost the deraining performance. Quantitative and qualitative experimental results demonstrate that the proposed method outperforms the state-of-the-art deraining methods. The source code will be available at \url{https://supercong94.wixsite.com/supercong94}.