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
Chenyang Gu, Mingyuan Zhang, Haozhe Xie +3 authors
A three-stage motion generation framework combines discrete token-based planning with diffusion-based synthesis to improve controllability and fidelity while reducing token usage and computational requirements.
Jiaxuan Lu, Kong Wang, Yemin Wang +9 authors
FinToolBench presents the first real-world benchmark for evaluating financial tool learning agents, featuring 760 executable tools and comprehensive evaluation criteria beyond simple execution success.
Ziyin Zhang, Zihan Liao, Hang Yu +2 authors
F2LLM-v2 is a multilingual embedding model family trained on 60 million samples across 200+ languages, achieving superior performance through LLM-based training, matryoshka learning, pruning, and distillation techniques.
Tiedong Liu, Wee Sun Lee
Vision-Language Models exhibit poor performance on visual entity tracking due to reliance on static features; a proposed method using spatiotemporal grounded chain-of-thought achieves high accuracy by generating object trajectories as intermediate states.
Lorenzo Mur-Labadia, Matthew Muckley, Amir Bar +6 authors
V-JEPA 2.1 is a self-supervised model that learns dense visual representations for images and videos through a combination of dense predictive loss, deep self-supervision, multi-modal tokenizers, and effective scaling.
Yichen Zhang, Da Peng, Zonghao Guo +19 authors
Cheers is a unified multimodal model that decouples visual details from semantic representations using a vision tokenizer, LLM-based Transformer, and cascaded flow matching head to achieve efficient joint optimization for both visual understanding and generation tasks.
Haitian Li, Haozhe Xie, Junxiang Xu +3 authors
MonoArt presents a unified framework for reconstructing articulated 3D objects from single images through progressive structural reasoning that enables stable articulation inference without external templates or multi-stage processes.
Dilxat Muhtar, Jiashun Liu, Wei Gao +8 authors
Complementary RL enables efficient agent learning by synchronizing experience extraction with policy optimization through dual objectives that evolve together during training.
Yuri Kuratov, Matvey Kairov, Aydar Bulatov +2 authors
GradMem enables efficient context storage and retrieval in language models through gradient-based memory writing that outperforms traditional forward-only approaches.
Hainuo Wang, Mingjia Li, Xiaojie Guo
Waypoint Diffusion Transformers address trajectory conflicts in pixel-space flow matching by using semantic waypoints from pre-trained vision models to disentangle generation paths and accelerate training convergence.
Lukas Hauzenberger, Niklas Schmidinger, Thomas Schmied +7 authors
There have been numerous attempts to distill quadratic attention-based large language models (LLMs) into sub-quadratic linearized architectures. However, despite extensive research, such distilled models often fail to match the performance of their teacher LLMs on various downstream tasks. We set out the goal of lossless distillation, which we define in terms of tolerance-corrected Win-and-Tie rates between student and teacher on sets of tasks. To this end, we introduce an effective distillation pipeline for xLSTM-based students. We propose an additional merging stage, where individually linearized experts are combined into a single model. We show the effectiveness of this pipeline by distilling base and instruction-tuned models from the Llama, Qwen, and Olmo families. In many settings, our xLSTM-based students recover most of the teacher's performance, and even exceed it on some downstream tasks. Our contributions are an important step towards more energy-efficient and cost-effective replacements for transformer-based LLMs.
Yuqing Wang, Chuofan Ma, Zhijie Lin +7 authors
CubiD is a discrete generation model for high-dimensional representations that enables fine-grained masking and learns rich correlations across spatial positions while maintaining fixed generation steps regardless of feature dimensionality.
Peng Sun, Jun Xie, Tao Lin
A data-efficient training framework for unified multimodal models that uses image-only data for pre-training followed by fine-tuning with mixed data types achieves state-of-the-art performance with reduced computational requirements.
Niladri Shekhar Dutt, Zifan Shi, Paul Guerrero +4 authors
Level-of-Semantics Tokenization (LoST) improves 3D shape generation by ordering tokens based on semantic salience and using a novel relational alignment loss for better reconstruction and efficiency.
Ming Li, Xirui Li, Tianyi Zhou
Large language models demonstrate varying capabilities in reasoning about unfolding geopolitical conflicts, showing strategic realism in structured settings but inconsistent performance in complex political environments.
Yiran Guan, Liang Yin, Dingkang Liang +5 authors
Video Streaming Thinking (VST) introduces a novel streaming video understanding paradigm that enables real-time reasoning during video playback through causal streaming adaptation and multi-turn interaction optimization.
Yaofeng Su, Yuming Li, Zeyue Xue +7 authors
OmniForcing distills a dual-stream bidirectional diffusion model into a streaming autoregressive generator while addressing training instability and synchronization issues through asymmetric alignment and specialized token mechanisms.
Dayuan Fu, Shenyu Wu, Yunze Wu +11 authors
OpenSWE presents the largest transparent framework for software engineering agent training, featuring 45,320 executable environments and achieving state-of-the-art performance on SWE-bench Verified.
Keda Tao, Yuhua Zheng, Jia Xu +13 authors
Long-form audio-visual comprehension benchmark reveals significant challenges for current omnimodal large language models in handling extended multi-modal inputs.
Hyungjoo Chae, Jungsoo Park, Alan Ritter
VeriEnv enables safe and scalable training of web agents by creating synthetic, verifiable environments from real websites through language model-based cloning.
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