Explore · 每周精选
发现最受关注的研究论文,追踪研究趋势,订阅感兴趣的期刊与关键词。
50 篇论文 · 按点赞排序
Jiawei Gu, Yunzhuo Hao, Huichen Will Wang +5 authors
ThinkMorph, a unified model fine-tuned on interleaved reasoning traces, enhances multimodal reasoning by generating coherent text-image steps, achieving significant performance gains and demonstrating emergent capabilities.
Ling-Team, Ang Li, Ben Liu +139 authors
Ling 2.0, a reasoning-oriented language model series, achieves high efficiency and accuracy through a Mixture-of-Experts paradigm, sparse activation, and innovative training techniques.
Zhaorun Chen, Zhuokai Zhao, Kai Zhang +15 authors
DreamGym is a unified framework that synthesizes diverse experiences for scalable online RL training, improving agent performance and reducing real-world interactions.
PAN Team, Jiannan Xiang, Yi Gu +31 authors
PAN, a general, interactable, and long-horizon world model, predicts future world states using a Generative Latent Prediction (GLP) architecture that combines autoregressive latent dynamics with a video diffusion decoder, enabling detailed, long-term, and coherent video simulation.
Mengzhao Chen, Meng Wu, Hui Jin +10 authors
A comprehensive comparison of low-precision floating-point and integer quantization in large language models reveals that fine-grained integer formats, especially MXINT8, offer superior accuracy and efficiency over floating-point formats in many cases.
Guoxin Chen, Zile Qiao, Xuanzhong Chen +13 authors
IterResearch, an iterative deep-research paradigm, improves long-horizon reasoning by reformulating it as a Markov Decision Process with strategic workspace reconstruction and Efficiency-Aware Policy Optimization, achieving better performance and interaction scaling compared to existing agents.
Artem Chervyakov, Ulyana Isaeva, Anton Emelyanov +15 authors
Mera Multi is an open multimodal evaluation framework for Russian-spoken architectures, addressing the lack of such benchmarks with 18 newly constructed tasks and a methodology to prevent benchmark leakage.
Cheng Yang, Haiyuan Wan, Yiran Peng +8 authors
VR-Bench evaluates video models' spatial reasoning capabilities through maze-solving tasks, demonstrating that these models excel in spatial perception and reasoning, outperforming VLMs and benefiting from diverse sampling during inference.
Chenze Shao, Darren Li, Fandong Meng +1 authors
Continuous Autoregressive Language Models (CALM) improve language model efficiency by predicting continuous vectors instead of discrete tokens, reducing computational cost while maintaining performance.
Mohammad Zbib, Hasan Abed Al Kader Hammoud, Sina Mukalled +5 authors
AraLingBench evaluates Arabic and bilingual LLMs' linguistic competence using a benchmark with expert-designed questions across grammar, morphology, spelling, reading comprehension, and syntax, revealing gaps between surface proficiency and true comprehension.
Qiushi Sun, Mukai Li, Zhoumianze Liu +11 authors
A hybrid safety detection framework combining formal verification and VLM-based contextual assessment improves the detection of unsafe operations in mobile agents.
Chunshi Wang, Junliang Ye, Yunhan Yang +6 authors
We introduce Part-X-MLLM, a native 3D multimodal large language model that unifies diverse 3D tasks by formulating them as programs in a structured, executable grammar. Given an RGB point cloud and a natural language prompt, our model autoregressively generates a single, coherent token sequence encoding part-level bounding boxes, semantic descriptions, and edit commands. This structured output serves as a versatile interface to drive downstream geometry-aware modules for part-based generation and editing. By decoupling the symbolic planning from the geometric synthesis, our approach allows any compatible geometry engine to be controlled through a single, language-native frontend. We pre-train a dual-encoder architecture to disentangle structure from semantics and instruction-tune the model on a large-scale, part-centric dataset. Experiments demonstrate that our model excels at producing high-quality, structured plans, enabling state-of-the-art performance in grounded Q\&A, compositional generation, and localized editing through one unified interface. Project page: https://chunshi.wang/Part-X-MLLM/
Ye Tian, Ling Yang, Jiongfan Yang +10 authors
A parallel multimodal diffusion framework, MMaDA-Parallel, enhances cross-modal alignment and semantic consistency in thinking-aware image synthesis by addressing error propagation issues in sequential approaches.
Tianhong Li, Kaiming He
Today's denoising diffusion models do not "denoise" in the classical sense, i.e., they do not directly predict clean images. Rather, the neural networks predict noise or a noised quantity. In this paper, we suggest that predicting clean data and predicting noised quantities are fundamentally different. According to the manifold assumption, natural data should lie on a low-dimensional manifold, whereas noised quantities do not. With this assumption, we advocate for models that directly predict clean data, which allows apparently under-capacity networks to operate effectively in very high-dimensional spaces. We show that simple, large-patch Transformers on pixels can be strong generative models: using no tokenizer, no pre-training, and no extra loss. Our approach is conceptually nothing more than "Just image Transformers", or JiT, as we call it. We report competitive results using JiT with large patch sizes of 16 and 32 on ImageNet at resolutions of 256 and 512, where predicting high-dimensional noised quantities can fail catastrophically. With our networks mapping back to the basics of the manifold, our research goes back to basics and pursues a self-contained paradigm for Transformer-based diffusion on raw natural data.
Zehong Ma, Longhui Wei, Shuai Wang +2 authors
The frequency-DeCoupled pixel diffusion framework improves image generation efficiency and quality by separating high-frequency details and low-frequency semantics, achieving superior performance compared to existing pixel diffusion models.
Kang Chen, Zhihao Liu, Tonghe Zhang +10 authors
The framework π<sub>RL</sub> uses reinforcement learning to train flow-based Vision-Language-Action models, addressing challenges with intractable action log-likelihoods and achieving significant performance improvements over supervised fine-tuning.
Rulin Shao, Akari Asai, Shannon Zejiang Shen +18 authors
Reinforcement Learning with Evolving Rubrics (RLER) enables training of deep research models for long-form tasks, outperforming existing models and proprietary systems while being more cost-effective.
Huijie Liu, Shuhao Cui, Haoxiang Cao +3 authors
A novel method, CoTyle, generates images in consistent visual styles using unique numerical style codes, filling an academic gap in code-to-style image generation.
Fei Tian, Xiangyu Tony Zhang, Yuxin Zhang +14 authors
Step-Audio-R1, using the Modality-Grounded Reasoning Distillation framework, achieves strong reasoning capabilities in audio, outperforming previous models and demonstrating the transferability of reasoning across modalities.
Wei He, Kai Han, Hang Zhou +4 authors
ROOT, a robust optimizer, enhances training stability and convergence for large language models by addressing dimensional fragility and outlier noise through adaptive Newton iterations and proximal optimization.
北京市昌平区探索星信息技术及软件开发工作室
京ICP备2026059466号