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。
Fan Jiang, Zhaoxu Sun, Mengchao Wang +38 authors
ABot-World-0 is a real-time action-conditioned video world model that uses progressive distillation, long-horizon alignment, and a co-designed streaming stack to enable efficient, long-horizon interactive world generation.
50 篇论文 · 按点赞排序
Kehan Li, Bohan Hou, Minghao Zhu +27 authors
RynnBrain 1.1 is a family of embodied foundation models that integrates spatio-temporal and physical grounding with contact-point prediction, 3D grounding, and cross-embodiment action spaces to improve robot perception, reasoning, and manipulation.
Yuhan Zhu, Changlian Ma, Xiangyu Zeng +12 authors
TimeLens2 improves video temporal grounding by treating evidence as interval sets with robust multi-span supervision and a Wasserstein-based reward, achieving strong results across benchmarks.
Shuqi Lu, Chaofan Li, Kun Luo +21 authors
AREX is a recursively self-improving deep research agent that verifies answers constraint-wise, compresses verified evidence into a compact state, and uses targeted follow-up research to refine results over long horizons.
Mikhail Komarov, Ivan Bondarenko, Stanislav Shtuka +5 authors
RAGU is a modular GraphRAG engine that separates knowledge extraction from consolidation and uses a small language-optimized model to build cleaner graphs, improving retrieval and synthesis over existing systems.
Runming He, Zhen Hao Wong, Hao Liang +4 authors
DataFlow-Harness guides LLM agents to build persistent, editable data pipelines via incremental mutations and live platform grounding, achieving high reliability with lower cost and latency.
Yijia Fan, Zonglin Di, Zimo Wen +8 authors
RESOURCE2SKILL converts diverse multimedia tutorials and artifacts into structured, executable skills for software agents, improving performance through hierarchical retrieval, composition, and online acquisition.
Xinyu Geng, Xuanhua He, Sixiang Chen +7 authors
Training tool-use agents to improve from their own experience remains challenging, as supervised fine-tuning relies on fixed teacher-distilled trajectories, while sparse-reward reinforcement learning provides weak supervision for long-horizon interactions. We present DeepSearch-Evolve, a self-distillation framework for web agents built on DeepSearch-World, a deterministic and verifiable environment with reproducible search and page-reading tools. DeepSearch-World contains 420K multi-hop QA tasks constructed from entity-level random walks and supports key agentic cognitive behaviors useful for self-evolving, including progress verification, grounded reflection, and failure recovery. DeepSearch-Evolve iteratively performs trajectory generation, filtering, data mixing, and fine-tuning to train stronger agents. Without distillation from more capable models, DeepSearch-World-9B achieves competitive performance compared with open-source agents, reaching 31.2% on BrowseComp, 61.5% on GAIA, and 93.4% on HotpotQA, showing that verifiable environments enable scalable self-evolution for long-horizon web agents. We will release the environment, 420K training pool, validation set, model, and code to facilitate future research on self-improving deep search agents.
Qing Zong, Yue Guo, Mengxin Yang +2 authors
EvolvingWorld is a framework and benchmark that models interactive literary worlds as long-horizon co-evolution processes using coupled character and world modules with LLM-based state maintenance and evaluation.
Guixu Lin, Zheng-Hui Huang, Siqi Yang +3 authors
AlayaRenderer-Flash accelerates a generative world renderer to real-time speeds via few-step autoregressive streaming and distilled codecs while preserving structured scene dynamics.
Yuhang Wang, Yuling Shi, Shaoqiu Zhang +6 authors
SWE-Pruner Pro uses an agent's internal representations to prune tool outputs directly, reducing token usage while maintaining or improving coding task performance.
Zitian Gao, Yilong Chen, Yihao Xiao +4 authors
Loopie is a looped Mixture-of-Experts Transformer that outperforms larger vanilla models at equal compute and achieves gold-medal reasoning on 2025 IMO and IPhO.
Xinjie Zhang, Peng Zhang, Shicheng Zheng +21 authors
Mage-Flow is a compact 4B generative stack combining a lightweight latent tokenizer and a native-resolution diffusion transformer to enable efficient high-resolution text-to-image generation and instruction-based editing.
Dongfang Li, Xiaodong Luo, Ruoyu Sun +62 authors
Researchers optimized trillion-parameter MoE post-training on Ascend NPUs to achieve high efficiency, then built specialized reasoning models for operations research via curated pre-training and supervised fine-tuning.
Maohua Li, Qirui Li, Yanke Zhou +10 authors
Text-to-image diffusion transformers use structural template tokens as implicit semantic registers that maintain object identity via indirect read-back from image latents, enabling training-free pruning of prompt-attending heads.
Xiaomi Robotics Team, Jun Guo, Piaopiao Jin +31 authors
Xiaomi-Robotics-1 is a vision-language-action model trained on extensive real-world manipulation data that performs diverse mobile tasks in unseen settings and adapts efficiently to new tasks with minimal fine-tuning.
Zishuo Li, Bowen Yang, Changtao Miao +29 authors
Open-AoE is an open egocentric manipulation dataset and toolchain that integrates scalable smartphone capture, structured processing, and downstream training support for embodied models.
Hao Liang, Qihan Lin, Zhaoyang Han +5 authors
A new benchmark and knowledge graph evaluate large language models on structured K-12 curriculum understanding, showing that graph-guided supervised fine-tuning improves performance across text and vision tasks.
Yiyang Cai, Nan Chen, Rongchang Xie +8 authors
HOMIE is a unified framework for human-object video personalization that improves interaction fidelity and intra-subject reference alignment via multimodal guidance and reference embeddings.
AlayaWorld Team, Kaipeng Zhang, Chuanhao Li +15 authors
AlayaWorld is an interactive long-horizon video world model that generates high-resolution video with persistent memory and efficient few-step inference for customizable virtual environments.
Xiangdong Zhang, Xiaohan Qin, Sunan Zou +10 authors
xHC enables large-scale residual-stream expansion in transformers via sparse updates and temporal feature augmentation, improving efficiency and downstream performance.
actAVA AI, Haolin Chen, Leon Qi +8 authors
Cura 1T is a healthcare-specialized LLM trained via recursive self-improvement to jointly handle consultation, clinical reasoning, diagnosis, and EHR tool use without degrading other capabilities.
Hanjing Ye, Tianle Zeng, Jiazhao Zhang +6 authors
ReferTrack improves embodied visual tracking by selecting image-grounded bounding boxes and decoding waypoints with temporal geometric tokens, achieving strong single-camera results.
Dingyun Zhang, Lixue Gong, Wei Liu
A unified model integrates video and image generation and editing by synthesizing video training data from image pairs via temporal flow warping, aligning modalities through mimic losses, and internalizing instruction-based region localization via sense-related tasks and latent and attention losses.
Yijun Liang, Yunjie Tian, Yijiang Li +4 authors
Visual Contrastive Self-Distillation improves on-policy self-distillation by contrasting original and content-erased image conditions to generate stronger training signals without external teachers or privileged data.
Runmao Yao, Kairui Hu, Yukang Cao +11 authors
Apple-PI introduces a physics-law-anchored benchmark with staged reasoning evaluation to diagnose where video generation models fail in simulating classical mechanics.
Junsong Chen, Jincheng Yu, Yitong Li +11 authors
SANA-Video 2.0 is a hybrid video diffusion transformer that combines linear and softmax attention to generate high-resolution video efficiently on a single GPU.
Maximo Eduardo Rulli, Thomas Vaitses Fontanari, Simone Petruzzi +9 authors
Diffusion language models encode diffusion timestep progress in residual streams, enabling extraction via probes and steering to modulate confidence and entropy.
Byeongho Heo, Jaehui Hwang, Sangdoo Yun +1 authors
On-policy delta distillation improves reasoning transfer by using the difference between a tuned teacher and its base model as a reward signal, yielding stronger performance with brief post-training.
Xiaokang Ma, Yifan Sun, Zhihong Jin +6 authors
ReflectWorld-MM is an entity-oriented multimodal memory system for open-ended video streams that organizes observations around persistent entities using hierarchical long-term memory to improve tracking and reasoning over time.
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