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
Zhen Yang, Wenyi Hong, Mingde Xu +5 authors
UI2Code$^\text{N}$, a visual language model enhanced through staged pretraining, fine-tuning, and reinforcement learning, achieves superior performance in UI-to-code generation, editing, and polishing with iterative feedback.
Jingxuan Wei, Caijun Jia, Xi Bai +7 authors
GGBench is introduced to evaluate geometric generative reasoning, addressing the gap in assessing integrated cognitive processes in multimodal models.
Dor Shmilovich, Tony Wu, Aviad Dahan +1 authors
LiteAttention exploits temporal coherence in diffusion attention to accelerate video generation without quality loss.
Junhao Cheng, Liang Hou, Xin Tao +1 authors
VANS, a model combining reinforcement learning, a Vision-Language Model, and a Video Diffusion Model, achieves state-of-the-art performance in Video-Next-Event Prediction by generating visually consistent and semantically accurate videos.
Ali Taghibakhshi, Sharath Turuvekere Sreenivas, Saurav Muralidharan +13 authors
Nemotron Elastic reduces training costs and memory usage by embedding multiple submodels within a single large language model, optimized for various deployment configurations and budgets without additional training or fine-tuning.
Jingtong Yue, Ziqi Huang, Zhaoxi Chen +3 authors
Video generation is evolving towards foundation models that integrate world simulation and rendering to produce physically plausible and interactive videos.
Worawalan Chatlatanagulchai, Hao Li, Yutaro Kashiwa +8 authors
Agentic coding tools receive goals written in natural language as input, break them down into specific tasks, and write or execute the actual code with minimal human intervention. Central to this process are agent context files ("READMEs for agents") that provide persistent, project-level instructions. In this paper, we conduct the first large-scale empirical study of 2,303 agent context files from 1,925 repositories to characterize their structure, maintenance, and content. We find that these files are not static documentation but complex, difficult-to-read artifacts that evolve like configuration code, maintained through frequent, small additions. Our content analysis of 16 instruction types shows that developers prioritize functional context, such as build and run commands (62.3%), implementation details (69.9%), and architecture (67.7%). We also identify a significant gap: non-functional requirements like security (14.5%) and performance (14.5%) are rarely specified. These findings indicate that while developers use context files to make agents functional, they provide few guardrails to ensure that agent-written code is secure or performant, highlighting the need for improved tooling and practices.
Xiaoshuai Hao, Lei Zhou, Zhijian Huang +41 authors
MiMo-Embodied, a cross-embodied foundation model, achieves state-of-the-art performance in both autonomous driving and embodied AI through multi-stage learning, curated data, and CoT/RL fine-tuning.
Geon Choi, Hangyul Yoon, Hyunju Shin +4 authors
A new instruction-guided lesion segmentation paradigm using a large-scale dataset and a vision-language model enables diverse CXR lesion segmentation with simple instructions.
Huiyi Chen, Jiawei Peng, Dehai Min +5 authors
MVI-Bench evaluates the robustness of Large Vision-Language Models against misleading visual inputs using a hierarchical taxonomy and a novel sensitivity metric, revealing significant vulnerabilities.
Jiaze Li, Hao Yin, Wenhui Tan +7 authors
Self-reflection mechanisms that rely on purely text-based rethinking processes perform well in most multimodal tasks. However, when directly applied to long-form video understanding scenarios, they exhibit clear limitations. The fundamental reasons for this lie in two points: (1)long-form video understanding involves richer and more dynamic visual input, meaning rethinking only the text information is insufficient and necessitates a further rethinking process specifically targeting visual information; (2) purely text-based reflection mechanisms lack cross-modal interaction capabilities, preventing them from fully integrating visual information during reflection. Motivated by these insights, we propose REVISOR (REflective VIsual Segment Oriented Reasoning), a novel framework for tool-augmented multimodal reflection. REVISOR enables MLLMs to collaboratively construct introspective reflection processes across textual and visual modalities, significantly enhancing their reasoning capability for long-form video understanding. To ensure that REVISOR can learn to accurately review video segments highly relevant to the question during reinforcement learning, we designed the Dual Attribution Decoupled Reward (DADR) mechanism. Integrated into the GRPO training strategy, this mechanism enforces causal alignment between the model's reasoning and the selected video evidence. Notably, the REVISOR framework significantly enhances long-form video understanding capability of MLLMs without requiring supplementary supervised fine-tuning or external models, achieving impressive results on four benchmarks including VideoMME, LongVideoBench, MLVU, and LVBench.
Senyu Fei, Siyin Wang, Li Ji +7 authors
Self-Referential Policy Optimization (SRPO) uses latent world representations to assign progress-wise rewards to failed trajectories, improving efficiency and effectiveness in vision-language-action robotic manipulation tasks without external demonstrations.
Yuqi Yin, Yibo Fu, Siyuan Wang +8 authors
Axonopedia, an AI agent utilizing LLMs and a multimodal foundation model, enhances property prediction and molecular design for Ionic Liquids through hierarchical search and real-world validation.
Özay Ezerceli, Mahmoud El Hussieni, Selva Taş +4 authors
TurkColBERT evaluates dense encoders and late-interaction models for Turkish information retrieval, demonstrating parameter efficiency and superior performance of late-interaction models with faster indexing.
Mingyue Cheng, Jie Ouyang, Shuo Yu +6 authors
A new training framework for RL-based LLM Agents is introduced, extending MDP methodology and demonstrating effectiveness on Multihop QA tasks.
N Dinesh Reddy, Sudeep Pillai
Orion, a visual agent framework, uses a suite of specialized computer vision tools to execute complex visual workflows, achieving competitive performance on multiple benchmarks and enabling autonomous visual reasoning.
Keda Tao, Kele Shao, Bohan Yu +3 authors
OmniZip is a training-free framework that compresses audio-visual tokens by dynamically pruning video tokens based on audio retention scores, achieving significant inference speedup and memory reduction without sacrificing performance.
Chaoyun Zhang, Liqun Li, He Huang +8 authors
UFO$^3$ unifies heterogeneous devices into a single orchestration fabric, enabling seamless task collaboration and dynamic optimization across distributed environments.
Junfu Pu, Teng Wang, Yixiao Ge +3 authors
ARC-Chapter is a large-scale video chaptering model that improves performance through extensive training data and a new evaluation metric, demonstrating superior results and transferability.
Hongwei Liu, Junnan Liu, Shudong Liu +33 authors
ATLAS, a large-scale, cross-disciplinary evaluation suite, addresses the limitations of existing benchmarks by providing high-difficulty, original, and high-fidelity scientific problems to assess the reasoning capabilities of Large Language Models.
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