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。
Dong Guo, Faming Wu, Feida Zhu +194 authors
Seed1.5-VL, a vision-language foundation model combining a vision encoder and a large MoE LLM, achieves state-of-the-art performance across various benchmarks and excels in multimodal reasoning tasks such as visual puzzles.
50 篇论文 · 按点赞排序
Bowen Zhang, Congchao Guo, Geng Yang +17 authors
MiniMax-Speech, an autoregressive Transformer-based TTS model, generates high-quality speech with a learnable speaker encoder that extracts reference speaker features without transcription, achieving SOTA results in voice cloning and supporting various extensions.
Zhiyuan Hu, Yibo Wang, Hanze Dong +5 authors
Explicit alignment of large reasoning models with deduction, induction, and abduction through a three-stage pipeline improves scalability and reliability in reasoning tasks.
Jiuhai Chen, Zhiyang Xu, Xichen Pan +10 authors
A diffusion transformer is used in a unified multimodal model framework to improve image generation while maintaining image understanding capabilities.
Xiaomi LLM-Core Team, Bingquan Xia, Bowen Shen +61 authors
MiMo-7B, a large language model optimized for reasoning tasks, enhances pre-training with data mixing and Multi-Token Prediction, and post-training with reinforcement learning on math and programming problems, achieving superior performance over larger models.
Mouxiang Chen, Binyuan Hui, Zeyu Cui +5 authors
Parallel scaling (ParScale) improves inference efficiency by reusing existing parameters and executing multiple transformations in parallel, offering superior performance with reduced memory and latency compared to parameter scaling.
Chenggang Zhao, Chengqi Deng, Chong Ruan +12 authors
DeepSeek-V3 addresses hardware limitations through MLA, MoE, FP8 training, and Multi-Plane Network Topology, enabling efficient large-scale LLM training and inference.
Yumin Choi, Jinheon Baek, Sung Ju Hwang
A meta-learning framework for optimizing system prompts in Large Language Models (LLMs) improves generalization across diverse tasks and datasets.
Weiyu Li, Xuanyang Zhang, Zheng Sun +15 authors
Step1X-3D addresses challenges in 3D generation through a hybrid VAE-DiT geometry generator and diffusion-based texture synthesis, achieving state-of-the-art results with an open-source framework.
Krzysztof Ociepa, Łukasz Flis, Remigiusz Kinas +2 authors
Bielik v3, a series of parameter-efficient generative text models, achieves high performance in Polish language processing with a custom tokenizer, Weighted Instruction Cross-Entropy Loss, and Adaptive Learning Rate.
Ke Wang, Junting Pan, Linda Wei +8 authors
A new image-to-code model and dataset are developed to enhance multimodal mathematical reasoning, resulting in a model that outperforms current benchmarks in mathematical problem-solving.
Krzysztof Ociepa, Łukasz Flis, Krzysztof Wróbel +2 authors
Bielik 11B v2, a scaled language model with 11B parameters, excels on Polish benchmarks through Weighted Instruction Cross-Entropy Loss and Adaptive Learning Rate, outperforming larger models and demonstrating resource-efficient deployment.
Tongxu Luo, Wenyu Du, Jiaxi Bi +5 authors
LeaP, a peer interaction mechanism for large reasoning models, improves error correction and performance across various math benchmarks by enabling collaborative reasoning paths.
Junjie Wang, Bin Chen, Yulin Li +3 authors
DeCLIP enhances CLIP by separating content and context features, improving performance on open-vocabulary dense prediction tasks like object detection and semantic segmentation.
Zhaochen Su, Linjie Li, Mingyang Song +8 authors
OpenThinkIMG and V-ToolRL enable LVLMs to learn adaptive policies for using external vision tools, outperforming existing methods on chart reasoning tasks.
Peng Sun, Yi Jiang, Tao Lin
A unified framework for continuous generative models improves training and sampling performance across multi-step and few-step methods.
Nadav Magar, Amir Hertz, Eric Tabellion +4 authors
A diffusion-based method fine-tuned on real and synthetic image pairs provides precise control over light sources and ambient illumination in images, offering more effective relighting than existing methods.
Binghai Wang, Runji Lin, Keming Lu +17 authors
World Preference Modeling (WorldPM) enhances preference fine-tuning through large-scale preference data, showing scalability benefits in adversarial and objective metrics, and improving generalization across various human preference datasets.
Zeyue Xue, Jie Wu, Yu Gao +8 authors
DanceGRPO is a unified RL framework that enhances visual generation across different paradigms, tasks, models, and reward systems, outperforming baselines in benchmarks and improving video generation stability.
Xiaokun Wang, Chris, Jiangbo Pei +9 authors
Skywork-VL Reward is a multimodal reward model that uses a large-scale preference dataset and Qwen2.5-VL-7B-Instruct architecture to achieve state-of-the-art performance in multimodal reasoning.
Aniruddha Roy, Pretam Ray, Abhilash Nandy +2 authors
Semi-automated frameworks using open-source small LLMs and reinforcement learning significantly improve instruction dataset generation for LLM fine-tuning across various tasks.
Kai Hua, Steven Wu, Ge Zhang +1 authors
AttentionInfluence, a training-free, unsupervised method using attention head masking, improves data selection for pretraining LLMs, enhancing their reasoning abilities across various benchmarks.
Bingxin Ke, Kevin Qu, Tianfu Wang +5 authors
Marigold, a family of conditional generative models, leverages pretrained latent diffusion models for dense image analysis tasks with minimal modification and minimal data.
Seongyun Lee, Seungone Kim, Minju Seo +9 authors
The CoT Encyclopedia is a bottom-up framework that automatically analyzes and categorizes reasoning strategies in long chain-of-thoughts generated by large language models, providing more interpretable analyses and enabling performance improvements.
Zachary Novack, Zach Evans, Zack Zukowski +8 authors
Adversarial Relativistic-Contrastive (ARC) post-training optimizes diffusion/flow models for fast text-to-audio generation with minimal latency.
Qingwen Bu, Yanting Yang, Jisong Cai +5 authors
UniVLA is a framework for learning cross-embodiment VLA policies using latent action models derived from internet-scale videos, achieving superior performance with reduced pretraining compute and downstream data.
Yuxin Jiang, Shengcong Chen, Siyuan Huang +8 authors
EnerVerse-AC, an action-conditional world model, enables realistic robotic inference and testing by simulating future actions and observations, thereby reducing costs and improving generalization in dynamic settings.
Chenxi Whitehouse, Tianlu Wang, Ping Yu +4 authors
A reinforcement learning approach called J1 improves the judgment ability of LLM-as-a-Judge models through verifiable rewards and chain-of-thought reasoning.
Wenxuan Wang, Fan Zhang, Yufeng Cui +5 authors
ETT is an end-to-end vision tokenizer tuning method that integrates visual tokenizer training with autoregressive tasks, significantly improving performance in multimodal understanding and visual generation.
Hu Yue, Siyuan Huang, Yue Liao +5 authors
A new benchmark framework evaluates text-to-video diffusion models in embodied AI for visual, motion, and semantic consistency, using a diverse dataset and evaluation toolkit.
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