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
Varun Singh, Lucas Krauss, Sami Jaghouar +23 authors
Arcee Trinity models are sparse Mixture-of-Experts architectures with varying parameter counts and activation patterns, utilizing advanced attention mechanisms and training optimizations.
Aniketh Garikaparthi, Manasi Patwardhan, Arman Cohan
ResearchGym presents a benchmark environment for evaluating AI agents on end-to-end research tasks, revealing significant capability-reliability gaps in current autonomous agents despite occasional state-of-the-art performance.
Saar Huberman, Kfir Goldberg, Or Patashnik +2 authors
Temporal statistics in semantic feature space provide a scalable approach for motion-centric video understanding, outperforming existing RGB, flow, and text-supervised methods.
Seonghyeon Ye, Yunhao Ge, Kaiyuan Zheng +33 authors
DreamZero is a World Action Model that leverages video diffusion to enable better generalization of physical motions across novel environments and embodiments compared to vision-language-action models.
Zikai Xiao, Jianhong Tu, Chuhang Zou +7 authors
WebWorld is an open-web simulator trained on over one million interactions that supports long-horizon reasoning and multi-format data, achieving performance comparable to advanced models like Gemini-3-Pro and GPT-4o.
Modi Jin, Yiming Zhang, Boyuan Sun +3 authors
GeoAgent achieves superior geolocation reasoning performance through a specialized dataset and reward mechanisms that ensure geographic accuracy and reasoning consistency.
Leon Liangyu Chen, Haoyu Ma, Zhipeng Fan +11 authors
UniT framework enables unified multimodal models to perform iterative reasoning and refinement through chain-of-thought test-time scaling, improving both generation and understanding capabilities.
Yiming Guan, Rui Yu, John Zhang +15 authors
A world model for desktop software that predicts UI state changes through textual description followed by visual synthesis, improving decision quality and execution robustness in computer-using tasks.
Yiwei Qin, Zhen Huang, Tiantian Mi +5 authors
Data Darwinism presents a systematic framework for data-model co-evolution through a ten-level taxonomy, demonstrating that advanced processing techniques significantly improve foundation model performance on scientific text.
Zun Li, John Schultz, Daniel Hennes +1 authors
AlphaEvolve, an evolutionary coding agent using large language models, automatically discovers new multiagent learning algorithms for imperfect-information games by evolving regret minimization and population-based training variants.
Shuofei Qiao, Yunxiang Wei, Xuehai Wang +10 authors
InnoEval is a deep innovation evaluation framework that emulates human-level idea assessment through knowledge-grounded, multi-perspective reasoning with heterogeneous deep knowledge search and multi-dimensional decoupled evaluation.
Xirui Li, Ming Li, Tianyi Zhou
Reinforcement learning (RL) with verifiable rewards has become a standard post-training stage for boosting visual reasoning in vision-language models, yet it remains unclear what capabilities RL actually improves compared with supervised fine-tuning as cold-start initialization (IN). End-to-end benchmark gains conflate multiple factors, making it difficult to attribute improvements to specific skills. To bridge the gap, we propose a Frankenstein-style analysis framework including: (i) functional localization via causal probing; (ii) update characterization via parameter comparison; and (iii) transferability test via model merging. Instead, RL induces a consistent inference-time shift primarily in mid-to-late layers, and these mid-to-late refinements are both transferable (via merging) and necessary (via freezing) for RL gains. Overall, our results suggest that RL's reliable contribution in visual reasoning is not a uniform enhancement of visual perception, but a systematic refinement of mid-to-late transformer computation that improves vision-to-reasoning alignment and reasoning performance, highlighting the limitations of benchmark-only evaluation for understanding multimodal reasoning improvements.
Wenxuan Ding, Nicholas Tomlin, Greg Durrett
Large language models can be improved for complex tasks by explicitly reasoning about cost-uncertainty tradeoffs through a Calibrate-Then-Act framework that enhances decision-making in sequential environments.
Stephan Rabanser, Sayash Kapoor, Peter Kirgis +3 authors
Traditional benchmark evaluations of AI agents fail to capture critical reliability issues, prompting the development of comprehensive metrics that assess consistency, robustness, predictability, and safety across multiple dimensions.
Nenad Tomašev, Matija Franklin, Simon Osindero
AI agents require adaptive frameworks for task decomposition and delegation that can dynamically respond to environmental changes and handle unexpected failures through structured authority transfer and trust mechanisms.
Shaobin Zhuang, Yuang Ai, Jiaming Han +12 authors
UniWeTok introduces a unified discrete tokenizer with a massive binary codebook and novel training techniques to achieve superior performance in image generation and multimodal tasks while reducing computational requirements.
Hee Seung Hwang, Xindi Wu, Sanghyuk Chun +1 authors
REFINE is a reinforcement learning framework that improves fast weight models for long-context modeling by training under next-sequence prediction instead of next-token prediction, enhancing their ability to capture long-range dependencies.
Muzhi Chen, Xuanhe Zhou, Wei Zhou +7 authors
This paper envisions a quantum database (Qute) that treats quantum computation as a first-class execution option. Unlike prior simulation-based methods that either run quantum algorithms on classical machines or adapt existing databases for quantum simulation, Qute instead (i) compiles an extended form of SQL into gate-efficient quantum circuits, (ii) employs a hybrid optimizer to dynamically select between quantum and classical execution plans, (iii) introduces selective quantum indexing, and (iv) designs fidelity-preserving storage to mitigate current qubit constraints. We also present a three-stage evolution roadmap toward quantum-native database. Finally, by deploying Qute on a real quantum processor (origin_wukong), we show that it outperforms a classical baseline at scale, and we release an open-source prototype at https://github.com/weAIDB/Qute.
Huanyao Zhang, Jiepeng Zhou, Bo Li +22 authors
A new benchmark called BrowseComp-V3 challenges multimodal large language models with complex, multi-hop reasoning tasks requiring deep search across text and visual modalities, revealing significant gaps in current capabilities.
Yandan Yang, Shuang Zeng, Tong Lin +11 authors
ABot-M0 presents a unified framework for embodied agent development that standardizes diverse robotic data and employs action manifold learning to improve prediction efficiency and stability.
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