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
Zihan Ma, Taolin Zhang, Maosong Cao +4 authors
A collaborative method combining human expertise and LLM reasoning enhances test-case generation for code evaluation, improving detection rates and verifier accuracy in benchmarks.
Sukjun Hwang, Brandon Wang, Albert Gu
Hierarchical networks (H-Nets) enable end-to-end learning by dynamically segmenting data, outperforming token-based models in various languages and modalities.
Eric Xing, Mingkai Deng, Jinyu Hou +1 authors
A new architecture for a general-purpose world model is proposed, based on hierarchical, multi-level, and mixed continuous/discrete representations, with a focus on simulating actionable possibilities for purposeful reasoning and acting.
Rongsheng Wang, Junying Chen, Ke Ji +4 authors
MedGen, a model trained on MedVideoCap-55K, achieves top performance in medical video generation, balancing visual quality and medical accuracy.
Yan Yang, Dongxu Li, Yutong Dai +11 authors
A GUI Test-time Scaling Agent addresses task planning ambiguity and visual grounding accuracy in GUI interactions using reinforcement learning and test-time scaling.
Dustin Wang, Rui-Jie Zhu, Steven Abreu +8 authors
Research evaluates various linear attention models in hybrid architectures, finding that selective gating, hierarchical recurrence, and controlled forgetting are crucial for effective recall in Transformers.
BAAI RoboBrain Team, Mingyu Cao, Huajie Tan +43 authors
RoboBrain 2.0, a heterogeneous vision-language model, excels in embodied reasoning tasks with strong performance on spatial and temporal benchmarks, supporting capabilities like spatial understanding and temporal decision-making.
Mohamed Elmoghany, Ryan Rossi, Seunghyun Yoon +26 authors
Research identifies key components and strategies for generating long-form videos with multiple characters, narrative coherence, and high-fidelity detail.
Xiangru Tang, Tianrui Qin, Tianhao Peng +15 authors
Agent KB, a hierarchical experience framework, enhances problem-solving success rates across different agents by enabling cross-agent knowledge transfer through a Reason-Retrieve-Refine pipeline.
Tianyu Zheng, Tianshun Xing, Qingshui Gu +10 authors
FR3E, a structured exploration framework, enhances LLM reasoning by providing targeted guidance at high-uncertainty decision points, leading to more stable training and accurate responses.
Guokan Shang, Hadi Abdine, Ahmad Chamma +7 authors
Nile-Chat models, using a Branch-Train-MiX strategy, outperform existing multilingual and Arabic LLMs on Egyptian dialect benchmarks, particularly in dual-script understanding and generation.
Zhiheng Xi, Guanyu Li, Yutao Fan +12 authors
A large-scale bilingual, multimodal, multi-disciplinary reasoning dataset (BMMR) is introduced to evaluate and develop large multimodal models (LMMs) across various disciplines and formats, with a focus on reasoning paths and discipline-specific performance.
Taolin Zhang, Zihan Ma, Maosong Cao +3 authors
The Code Triangle framework evaluates LLMs in code generation across editorial analysis, implementation, and test case generation, revealing areas for improvement through human-generated inputs and model mixtures.
François Rozet, Ruben Ohana, Michael McCabe +3 authors
Latent-space diffusion models effectively emulate dynamical systems with high accuracy and diversity, even at high compression rates, outperforming non-generative methods.
Modi Shi, Li Chen, Jin Chen +7 authors
Investigation into data diversity in robotic manipulation reveals that task diversity is more critical than demonstration quantity, single-embodiment data can suffice for cross-embodiment transfer, and expert diversity can be confounding, leading to a distribution debiasing method that improves performance.
Taekyung Kim, Dongyoon Han, Byeongho Heo +2 authors
ToBo, a self-supervised learning pipeline, generates compact and temporally aware visual representations by encoding scenes into a bottleneck token and predicting subsequent scenes with minimal hints, demonstrating superior performance in sequential tasks.
Nadav Schneider, Itamar Zimerman, Eliya Nachmani
A novel differential mechanism for Mamba, a selective state-space layer architecture, improves language modeling performance by addressing overallocation of attention to irrelevant context.
Baolong Bi, Shenghua Liu, Xingzhang Ren +7 authors
RefineX is a framework that uses programmatic editing tasks to efficiently refine large-scale pre-training data for LLMs, improving text quality and performance across various tasks.
Kaiqu Liang, Haimin Hu, Xuandong Zhao +3 authors
Research introduces machine bullshit as a framework to assess truthfulness in LLMs, using the Bullshit Index and taxonomy, and finds RLHF and CoT prompting increase specific bullshit forms.
Ramchalam Kinattinkara Ramakrishnan, Zhaocong Yuan, Shaojie Zhuo +4 authors
OmniDraft is a unified framework enabling a single draft model to dynamically adapt to various target models, addressing cross-vocabulary mismatches and improving decoding speed for on-device LLM applications.
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