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。
Weiyun Wang, Zhangwei Gao, Lixin Gu +58 authors
InternVL 3.5 introduces Cascade RL, ViR, and DvD to enhance reasoning, efficiency, and performance in multimodal models.
50 篇论文 · 按点赞排序
Zhiliang Peng, Jianwei Yu, Wenhui Wang +10 authors
VibeVoice synthesizes long-form multi-speaker speech using next-token diffusion and a highly efficient continuous speech tokenizer, achieving superior performance and fidelity.
Huichi Zhou, Yihang Chen, Siyuan Guo +8 authors
A novel memory-augmented reinforcement learning paradigm enables adaptive LLM agents to continually learn without fine-tuning, using episodic memory and a neural case-selection policy.
Ning Shang, Yifei Liu, Yi Zhu +12 authors
rStar2-Agent, a 14B math reasoning model trained with agentic reinforcement learning, achieves state-of-the-art performance by efficiently handling complex problem-solving with advanced cognitive behaviors and minimal computational resources.
Xiao Liang, Zhongzhi Li, Yeyun Gong +4 authors
An online self-play strategy with variational problem synthesis for RLVR training maintains policy entropy and improves Pass@k performance on reasoning benchmarks.
Neta Glazer, Yael Segal-Feldman, Hilit Segev +6 authors
Interpretability methods like logit lens, linear probing, and activation patching are applied to ASR to uncover internal dynamics, repetition hallucinations, and semantic biases, enhancing model transparency and robustness.
Yibin Wang, Zhimin Li, Yuhang Zang +6 authors
Pref-GRPO, a pairwise preference reward-based GRPO method, enhances text-to-image generation by mitigating reward hacking and improving stability, while UniGenBench provides a comprehensive benchmark for evaluating T2I models.
Zongxia Li, Wenhao Yu, Chengsong Huang +8 authors
Vision-SR1 uses reinforcement learning to enhance visual reasoning in vision-language models by decomposing the process into visual perception and language reasoning stages, improving accuracy and reducing hallucinations.
Yizhi Li, Qingshui Gu, Zhoufutu Wen +14 authors
TreePO, a self-guided rollout algorithm for sequence generation, reduces computational cost and enhances exploration diversity in reinforcement learning for large language models.
Dawei Gao, Zitao Li, Yuexiang Xie +20 authors
AgentScope enhances agentic applications by providing flexible tool-based interactions, unified interfaces, and advanced infrastructure based on the ReAct paradigm, supporting efficient and safe development and deployment.
Zhenting Wang, Qi Chang, Hemani Patel +8 authors
MCP-Bench evaluates large language models on complex, multi-step tasks requiring tool use, coordination, and planning across various domains.
Ryan Teknium, Roger Jin, Jai Suphavadeeprasit +6 authors
Hermes 4, a hybrid reasoning model, integrates structured multi-turn reasoning with broad instruction-following, evaluated across various benchmarks including math, coding, knowledge, comprehension, and alignment.
Shaojin Wu, Mengqi Huang, Yufeng Cheng +5 authors
USO, a unified model, achieves state-of-the-art performance in both style similarity and subject consistency by disentangling and re-composing content and style through a disentangled learning scheme and style reward-learning paradigm.
Weida Wang, Dongchen Huang, Jiatong Li +32 authors
CMPhysBench evaluates LLMs in condensed matter physics using calculation problems and a new SEED score for partial credit assessment, revealing significant capability gaps.
Kaijun Wang, Liqin Lu, Mingyu Liu +7 authors
ODYSSEY is a unified mobile manipulation framework for quadruped robots that integrates high-level task planning with low-level whole-body control, addressing challenges in egocentric perception, generalization, and coordination in unstructured environments.
Lin Li, Zehuan Huang, Haoran Feng +4 authors
VoxHammer is a training-free method that performs precise and coherent 3D editing in latent space, ensuring consistency in preserved regions and high-quality overall results.
Jianwen Jiang, Weihong Zeng, Zerong Zheng +6 authors
A framework using Multimodal Large Language Models and a specialized Multimodal DiT architecture generates semantically coherent and expressive character animations from multimodal inputs.
Yaqi Li, Peng Chen, Mingyang Han +6 authors
The Visual-Chain of Guidance (Visual-CoG) paradigm enhances text-to-image generation by providing stage-aware rewards, improving performance across multiple benchmarks.
Chengyue Yu, Siyuan Lu, Chenyi Zhuang +14 authors
AWorld, an open-source system for large-scale agent-environment interaction, accelerates experience collection and enhances reinforcement learning, leading to significant improvements in agentic AI performance on complex benchmarks.
Yosef Dayani, Omer Benishu, Sagie Benaim
MV-RAG enhances text-to-3D generation by retrieving 2D images and conditioning a multiview diffusion model to improve consistency and accuracy, especially for out-of-domain concepts.
Long Le, Ryan Lucas, Chen Wang +4 authors
PIXIE, a neural network method, predicts physical properties of 3D scenes from visual features, enabling fast and realistic physics simulation using supervised learning and pretrained visual features.
Zeyi Sun, Yuhang Cao, Jianze Liang +8 authors
CODA, a trainable compositional framework, combines a generalist planner and specialist executor to achieve robust execution and cross-domain generalization in scientific computing GUIs.
Zihao Huang, Yu Bao, Qiyang Min +8 authors
UltraMemV2, a redesigned memory-layer architecture, achieves performance parity with 8-expert MoE models while significantly reducing memory access costs.
Shengqu Cai, Ceyuan Yang, Lvmin Zhang +10 authors
Long video generation is addressed by introducing a sparse attention routing module, Mixture of Contexts, to efficiently manage long-term memory and retrieval in diffusion transformers.
Samuel Lewis-Lim, Xingwei Tan, Zhixue Zhao +1 authors
Investigation into Chain-of-Thought dynamics and faithfulness across various models reveals inconsistencies in their reliance on CoT and its alignment with actual reasoning.
Zhixuan Liang, Yizhuo Li, Tianshuo Yang +7 authors
Discrete Diffusion VLA uses a single-transformer policy with discrete diffusion to model actions, improving decoding order, consistency, and performance over autoregressive and continuous diffusion methods.
Heng Lin, Zhongwen Xu
Tool-Integrated Reasoning (TIR) enhances Large Language Models (LLMs) by expanding their problem-solving capabilities through the use of external tools, and Advantage Shaping Policy Optimization (ASPO) improves model behavior and tool usage.
Minhyeong Lee, Suyoung Hwang, Seunghyun Moon +13 authors
Spacer, a scientific discovery system, uses deliberate decontextualization to generate creative and factually grounded scientific concepts from keyword sets, achieving high accuracy and similarity to leading publications.
Miran Heo, Sukjun Hwang, Min-Hung Chen +4 authors
AUSM, an autoregressive universal segmentation model, unifies prompted and unprompted video segmentation by treating it as sequential mask prediction, achieving superior performance and faster training on standard benchmarks.
Ming Chen, Liyuan Cui, Wenyuan Zhang +5 authors
An autoregressive video generation framework with multimodal control and low-latency extrapolation uses a modified large language model and a deep compression autoencoder to achieve high efficiency and fine-grained controllability.
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