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
Jiajia Lin, Mingxuan Du, Tuowen Zhou +2 authors
A new benchmark and mesh-based evaluation reveal that multi-person image editors fail at anatomical consistency and contact geometry despite high VLM scores.
Haopeng Li, Yitong Li, Junsong Chen +8 authors
Sol-Attn improves training-free sparse attention for diffusion transformers by combining dynamic block routing, sparse computation, and approximation correction in a single online pass to accelerate video generation without sacrificing quality.
Senqiao Yang, Kaichen Zhang, Zhaoyang Jia +20 authors
Mage-VL is a codec-native streaming vision-language model that uses motion-aware tokenization and a dual-system architecture to enable efficient real-time multimodal perception and reasoning.
Yifan Ye, Yankai Fu, Yaoxu Lv +26 authors
The paper organizes embodied robot learning data into a scalable pyramid and analyzes how mixing sources affects foundation model capabilities and open challenges.
Yuhang Zhu, Mingxuan Du, Benfeng Xu +3 authors
PALATE introduces a scalable, user-simulated benchmark for role-playing agents that evaluates multi-turn dialogue quality and personalized satisfaction via tailored rubrics rather than fixed continuations.
Haolei Xu, Xiaowen Xu, Haiwen Hong +5 authors
Relay-OPD detects reasoning errors via teacher-student trajectory divergence and briefly delegates generation to the teacher to recover valid supervision, improving small-model math reasoning while cutting training cost.
Ruizhe Li, Mingxuan Du, Benfeng Xu +1 authors
InMind reveals that long-term memory systems fail to retrieve implicitly related facts, showing the bottleneck lies in query-conditioned retrieval rather than storage.
Jian Hu, Huiying Li, Hao Zhang +8 authors
Molt is a compact PyTorch framework for agentic reinforcement learning that enables efficient asynchronous training of multimodal and mixture-of-experts policies with minimal overhead.
Tengfei Liu, Yang Shi, Yuran Wang +16 authors
RefCaptioner is a two-stage framework for multi-reference image-grounded video captioning that improves phrase-level grounding and factual consistency, supported by a new benchmark and training corpus.
Zhiyuan Yao, Yuxin Chen, Zhengxi Lu +13 authors
SkillRise is a unified reinforcement learning framework that trains LLM agents to extract, refine, and reuse transferable skills across related tasks through progressive sequences and decoupled credit assignment.
Yihao Chen, Shi Chang, Khaled Chawa +4 authors
MindForge automates creation of source-free training environments from command-line programs to improve from-scratch code generation via curated synthesis trajectories.
Haoyuan Wu, Aoqi Wu, Hai Wang +3 authors
Native multimodal pre-training of vision-language transformers follows predictable compute scaling laws with distinct language and multimodal allocation behaviors, enabling efficient resource allocation and cross-modal transfer.
Yong Liu, Xiaolong Fu, Zihang Xu +10 authors
Oxygen-TryOn is a fashion-native foundation model that performs multi-reference virtual try-on across diverse categories and scenarios using a three-stage training recipe and hybrid reinforcement learning rewards.
Gaurav Dadhich
Agentic Context Management is proposed as a lifecycle discipline to actively control reasoning context, reducing quadratic token costs through structured ingestion, scoping, anticipation, and validated compaction.
Jun Zhan, Chen Yang, Yitian Gong +23 authors
OmniVAE jointly trains an audio-video VAE with contrastive and distillation objectives to align cross-modal latent spaces and improve synchronized generation.
Tianyi Men, Zhuoran Jin, Kang Liu +1 authors
Researchers propose a controlled multi-turn framework to study long-horizon planning acquisition, shaping via reinforcement learning, and integration across environments.
Yang Zhou, Zixuan Huang, Sunzhu Li +10 authors
SpatialCLI teaches vision-language models to use and internalize specialist spatial tools, improving embodied task performance with and without external perception aids.
Siyu Yan, Zhuoran Yan, Haiying Xu +10 authors
See2Think evaluates whether multimodal language models genuinely use intermediate visual reasoning states through a benchmark of visually dependent tasks and a process-tracing framework, revealing that faithful rendering is a key bottleneck and that models depend behaviorally on visual feedback.
Jiawei Xu, Minghui Liu, Juzheng Zhang +2 authors
β-OPSD generalizes on-policy self-distillation via a tunable KL-regularized policy optimization objective, using geometric interpolation of teacher and reference logits for stable, efficient reasoning model training.
Antonia Calvi, Avinash Sooriyarachchi, Giada Pistilli +8 authors
Shieldstral is a compact multimodal safety classifier that unifies diverse moderation tasks as binary questions to achieve high performance with far fewer parameters.
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