OS-ATLAS: A Foundation Action Model for Generalist GUI Agents
Zhiyong Wu, Zhenyu Wu, Fangzhi Xu +8 authors
OS-Atlas improves GUI agent performance through a large open-source GUI grounding dataset and model innovations.
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
Cut Your Losses in Large-Vocabulary Language Models
Erik Wijmans, Brody Huval, Alexander Hertzberg +2 authors
Cut Cross-Entropy reduces the memory footprint of large language models during training by efficiently computing cross-entropy loss without materializing logits for all tokens.
AndroidLab: Training and Systematic Benchmarking of Android Autonomous Agents
Yifan Xu, Xiao Liu, Xueqiao Sun +7 authors
AndroidLab provides a systematic framework for training and evaluating Android agents, supporting both large language models and multimodal models, and improves their task success rates.
Multimodal Autoregressive Pre-training of Large Vision Encoders
Enrico Fini, Mustafa Shukor, Xiujun Li +13 authors
AIMV2, a multimodal vision encoder paired with an autoregressive decoder, achieves superior performance in vision benchmarks and multimodal image understanding tasks compared to contrastive models.
BlueLM-V-3B: Algorithm and System Co-Design for Multimodal Large Language Models on Mobile Devices
Xudong Lu, Yinghao Chen, Cheng Chen +19 authors
BlueLM-V-3B optimizes the deployment of multimodal large language models on mobile platforms through redesigned dynamic resolution and hardware-aware system optimizations, achieving high performance with a small model size and fast generation speed.
O1 Replication Journey -- Part 2: Surpassing O1-preview through Simple Distillation, Big Progress or Bitter Lesson?
Zhen Huang, Haoyang Zou, Xuefeng Li +7 authors
Knowledge distillation from O1 model combined with fine-tuning achieves superior performance on complex mathematical reasoning and generalizes to open-ended QA tasks, promoting transparency in AI research.
M-Longdoc: A Benchmark For Multimodal Super-Long Document Understanding And A Retrieval-Aware Tuning Framework
Yew Ken Chia, Liying Cheng, Hou Pong Chan +5 authors
A benchmark and framework named M-LongDoc evaluates and improves large multimodal models for document reading, focusing on open-ended solutions and retrieval tasks over lengthy documents.
Critic-V: VLM Critics Help Catch VLM Errors in Multimodal Reasoning
Di Zhang, Jingdi Lei, Junxian Li +10 authors
The Critic-V framework leverages an Actor-Critic paradigm to enhance multimodal reasoning in vision-language models by decoupling reasoning and critique processes, significantly improving performance and reliability.
Style-Friendly SNR Sampler for Style-Driven Generation
Jooyoung Choi, Chaehun Shin, Yeongtak Oh +2 authors
The Style-friendly SNR sampler modifies the noise level distribution during fine-tuning to improve style alignment in diffusion models, enabling better capture of unique artistic styles.
From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge
Dawei Li, Bohan Jiang, Liangjie Huang +10 authors
A survey of LLM-based judgment and assessment explores definitions, taxonomy, benchmarks, challenges, and future directions for the LLM-as-a-judge paradigm.
Large-Scale Text-to-Image Model with Inpainting is a Zero-Shot Subject-Driven Image Generator
Chaehun Shin, Jooyoung Choi, Heeseung Kim +1 authors
Diptych Prompting is a zero-shot text-to-image generation method that enhances subject alignment and attention between panels during inpainting, resulting in visually preferable images across various applications.
GMAI-VL & GMAI-VL-5.5M: A Large Vision-Language Model and A Comprehensive Multimodal Dataset Towards General Medical AI
Tianbin Li, Yanzhou Su, Wei Li +15 authors
GMAI-VL, a vision-language model trained on the GMAI-VL-5.5M dataset, achieves top performance in multimodal medical tasks by integrating visual and textual information.
Stronger Models are NOT Stronger Teachers for Instruction Tuning
Zhangchen Xu, Fengqing Jiang, Luyao Niu +2 authors
The Larger Models' Paradox reveals that larger models are not always better teachers for fine-tuning smaller models, and a new metric, Compatibility-Adjusted Reward (CAR), is introduced to measure and improve the effectiveness of response generators.
LLM2CLIP: Powerful Language Model Unlock Richer Visual Representation
Weiquan Huang, Aoqi Wu, Yifan Yang +8 authors
LLM2CLIP leverages large language models to enhance CLIP's multimodal representation learning, enabling better handling of complex image captions by fine-tuning the LLM and using contrastive learning.
Identity-Preserving Text-to-Video Generation by Frequency Decomposition
Shenghai Yuan, Jinfa Huang, Xianyi He +5 authors
ConsisID is a tuning-free, DiT-based IPT2V model that uses frequency-aware identity control to generate high-quality, identity-preserving videos by decomposing facial features into low and high-frequency elements.
DynaSaur: Large Language Agents Beyond Predefined Actions
Dang Nguyen, Viet Dac Lai, Seunghyun Yoon +9 authors
A new LLM agent framework enables dynamic action creation and reuse by generating and executing programs, offering greater flexibility and improved performance in complex, real-world scenarios.
Pathways on the Image Manifold: Image Editing via Video Generation
Noam Rotstein, Gal Yona, Daniel Silver +3 authors
Image editing is improved using pretrained video models to ensure accurate edits and preserve image fidelity by reformulating the process as a smooth temporal transition.
Language Models are Hidden Reasoners: Unlocking Latent Reasoning Capabilities via Self-Rewarding
Haolin Chen, Yihao Feng, Zuxin Liu +8 authors
LaTent Reasoning Optimization (LaTRO) enhances LLMs' reasoning capabilities through variational optimization of latent distributions, improving zero-shot accuracy on complex reasoning tasks.
WebRL: Training LLM Web Agents via Self-Evolving Online Curriculum Reinforcement Learning
Zehan Qi, Xiao Liu, Iat Long Iong +10 authors
WebRL enhances open LLMs as web agents through a self-evolving curriculum, robust outcome-supervised reward model, and adaptive reinforcement learning strategies, achieving high success rates compared to proprietary models.
OpenScholar: Synthesizing Scientific Literature with Retrieval-augmented LMs
Akari Asai, Jacqueline He, Rulin Shao +22 authors
OpenScholar, a specialized language model, enhances scientific query synthesis by retrieving relevant passages and providing citation-backed answers, outperforming existing models in correctness and citation accuracy.