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。
Zhe Chen, Weiyun Wang, Yue Cao +37 authors
InternVL 2.5, an advanced multimodal large language model, showcases competitive performance across various benchmarks, including multimodal reasoning and understanding, and is the first open-source model to surpass 70% on the MMMU benchmark using Chain-of-Thought reasoning.
50 篇论文 · 按点赞排序
Marah Abdin, Jyoti Aneja, Harkirat Behl +24 authors
A 14-billion parameter language model surpasses its teacher model in STEM-focused QA through strategic use of synthetic data, improved data quality, and enhanced training techniques.
Pan Zhang, Xiaoyi Dong, Yuhang Cao +26 authors
A framework combining disentangled perception, reasoning, and memory mechanisms enables real-time interaction for long-term multimodal AI systems.
Shibo Hao, Sainbayar Sukhbaatar, DiJia Su +4 authors
A new reasoning paradigm, Coconut, utilizes continuous representations in latent space to enhance LLM performance on complex reasoning tasks, particularly those requiring backtracking.
Chujie Zheng, Zhenru Zhang, Beichen Zhang +6 authors
ProcessBench evaluates models' ability to identify errors in mathematical reasoning steps, showing that existing process reward models struggle with difficult problems and underperform compared to critic models and a fine-tuned PRM.
Zongyu Lin, Wei Liu, Chen Chen +14 authors
STIV, a text-image-conditioned video generation method integrating Diffusion Transformer and classifier-free guidance, achieves state-of-the-art performance in text-to-video, text-image-to-video, and image-to-video tasks.
Egor Cherepanov, Nikita Kachaev, Artem Zholus +2 authors
The paper provides a framework for understanding and evaluating different types of memory in reinforcement learning agents, using cognitive science-inspired definitions and a standardized experimental methodology.
Jiarui Zhang, Ollie Liu, Tianyu Yu +2 authors
A benchmark called Geoperception is introduced to evaluate MLLMs' geometric description accuracy, leading to the development of Euclid, a model optimized for low-level geometric perception using synthetic data and a data curriculum.
LG AI Research, Soyoung An, Kyunghoon Bae +30 authors
EXAONE 3.5 language models, available in three configurations, demonstrate exceptional instruction following, long-context comprehension, and competitive performance across various benchmarks.
Yutao Sun, Hangbo Bao, Wenhui Wang +5 authors
LatentLM integrates continuous and discrete data using causal Transformers, VAEs, and next-token diffusion, achieving superior performance in multimodal tasks like image generation, large language model integration, and text-to-speech synthesis.
Zhisheng Zhong, Chengyao Wang, Yuqi Liu +12 authors
As Multi-modal Large Language Models (MLLMs) evolve, expanding beyond single-domain capabilities is essential to meet the demands for more versatile and efficient AI. However, previous omni-models have insufficiently explored speech, neglecting its integration with multi-modality. We introduce Lyra, an efficient MLLM that enhances multimodal abilities, including advanced long-speech comprehension, sound understanding, cross-modality efficiency, and seamless speech interaction. To achieve efficiency and speech-centric capabilities, Lyra employs three strategies: (1) leveraging existing open-source large models and a proposed multi-modality LoRA to reduce training costs and data requirements; (2) using a latent multi-modality regularizer and extractor to strengthen the relationship between speech and other modalities, thereby enhancing model performance; and (3) constructing a high-quality, extensive dataset that includes 1.5M multi-modal (language, vision, audio) data samples and 12K long speech samples, enabling Lyra to handle complex long speech inputs and achieve more robust omni-cognition. Compared to other omni-methods, Lyra achieves state-of-the-art performance on various vision-language, vision-speech, and speech-language benchmarks, while also using fewer computational resources and less training data.
Jianzong Wu, Chao Tang, Jingbo Wang +3 authors
DiffSensei, a diffusion-based framework with a multimodal large language model, enhances manga generation by precisely controlling multi-character appearances and interactions using masked cross-attention and text cues.
Jian Yang, Jiaxi Yang, Ke Jin +7 authors
A human-curated benchmark (CodeArena) and a large synthetic instruction corpus (SynCode-Instruct) are introduced to evaluate code LLMs based on human preference alignment, revealing performance differences between open-source and proprietary models.
Yibin Wang, Zhiyu Tan, Junyan Wang +3 authors
LiFT fine-tunes text-to-video generative models using human feedback to improve video alignment with text descriptions, demonstrating superior performance over a larger model.
Jarvis Guo, Tuney Zheng, Yuelin Bai +7 authors
A large-scale multimodal instruction-tuning dataset with intermediate rationales improves reasoning capabilities in large language models, achieving state-of-the-art performance across benchmarks.
Zejian Li, Chenye Meng, Yize Li +9 authors
A new dataset with structural scene graph annotations improves text-to-image generation performance on complex scenes, benchmarked by CompSG-Bench.
Jianhong Bai, Menghan Xia, Xintao Wang +6 authors
A plug-and-play module enhances text-to-video models for multi-camera video generation, incorporating 6 DoF camera poses and ensuring dynamic consistency across viewpoints using a hybrid training scheme and multi-view synchronization.
Trong-Tung Nguyen, Quang Nguyen, Khoi Nguyen +2 authors
SwiftEdit performs instant text-guided image editing with high efficiency using a one-step inversion framework and mask-guided editing with attention rescaling.
Yuan Liu, Le Tian, Xiao Zhou +4 authors
POINTS1.5, an enhanced vision-language model, incorporates a dynamic high-resolution vision encoder, bilingual support, and rigorous dataset filtering methods to achieve superior performance.
Hanqing Zhu, Zhenyu Zhang, Wenyan Cong +7 authors
APOLLO, an approximated gradient scaling optimizer for LLMs, reduces memory usage without compromising performance, enhancing throughput and scalability.
Zijian Zhou, Shikun Liu, Xiao Han +11 authors
Leffa, a flow field learning method integrated into diffusion models, enhances controllable person image generation by addressing fine-grained detail distortions.
Dongting Hu, Jierun Chen, Xijie Huang +16 authors
SnapGen, a compact and fast text-to-image diffusion model, generates high-resolution images on mobile devices using design optimizations, knowledge distillation, and adversarial guidance.
Yiheng Xu, Dunjie Lu, Zhennan Shen +5 authors
AgentTrek synthesizes high-quality GUI agent trajectories using web tutorials, improving their performance cost-effectively compared to human annotation.
Kasra Arabi, Benjamin Feuer, R. Teal Witter +2 authors
The proposed two-stage watermarking framework using Fourier patterns and diffusion model's initial noise enhances robustness to forgery and removal in AI-generated images.
Jinyi Hu, Shengding Hu, Yuxuan Song +6 authors
ACDiT, an Autoregressive blockwise Conditional Diffusion Transformer, bridges visual generation and autoregressive modeling, demonstrating effectiveness in image and video generation and potential for visual understanding tasks.
Xi Chen, Zhifei Zhang, He Zhang +10 authors
UniReal treats image generation and editing tasks as discontinuous video generation to capture visual variations and consistency, learning from large-scale video data.
Nahid Alam, Karthik Reddy Kanjula, Surya Guthikonda +16 authors
Maya is an open-source multimodal multilingual model that addresses the limitations of Vision-Language Models in handling low-resource languages and cultural nuances by introducing a multilingual pretraining dataset and analyzing/removing toxicity.
Thibault Le Sellier De Chezelles, Maxime Gasse, Alexandre Drouin +17 authors
BrowserGym provides a standardized environment for evaluating web agents using LLMs, facilitating consistent benchmarking and improving agent development and analysis.
Nicolas Dufour, David Picard, Vicky Kalogeiton +1 authors
A generative geolocation approach using diffusion and Riemannian flow matching achieves state-of-the-art performance on visual geolocation tasks and introduces probabilistic geolocation with new metrics.
Yi Chen, Yuying Ge, Yizhuo Li +4 authors
A generative pre-training approach using motion token sequences from video data enhances robot learning, enabling transfer to real manipulation tasks with improved robustness and efficiency.
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