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
Neil Rathi, Alec Radford
Token filtering during pretraining effectively reduces unwanted language model capabilities while maintaining alignment, becoming more effective at larger scales and tolerating noisy labels with sufficient compute.
Bin Tan, Changjiang Sun, Xiage Qin +8 authors
LingBot-Depth is a depth completion model that uses visual context to refine depth maps through masked depth modeling and automated data curation for improved spatial perception in robotics and autonomous systems.
Dohun Lee, Chun-Hao Paul Huang, Xuelin Chen +3 authors
Memory-V2V enhances multi-turn video editing by maintaining cross-consistency through explicit memory mechanisms and efficient token compression in video-to-video diffusion models.
Chen Chen, Lai Wei
A novel Post-LayerNorm Transformer architecture called Keel addresses training instability in extremely deep networks by replacing residual connections with Highway-style connections, enabling stable training beyond 1000 layers.
Zhiliang Peng, Jianwei Yu, Yaoyao Chang +21 authors
VibeVoice-ASR is a unified end-to-end speech understanding framework that processes long-form audio without chunking, supports multiple languages and code-switching, and uses prompt-based context injection for improved domain-specific accuracy.
Jinyang Wu, Shuo Yang, Changpeng Yang +4 authors
Spark is a reinforcement learning framework that strategically allocates computational resources by branching at critical decision states, improving sample efficiency and generalization for long-horizon tasks.
Jialong Wu, Xiaoying Zhang, Hongyi Yuan +7 authors
Visual generation enhances reasoning capabilities in multimodal models by providing more natural world models for physical and spatial tasks, while verbal reasoning remains sufficient for abstract domains.
Sangwon Jang, Taekyung Ki, Jaehyeong Jo +3 authors
Self-refining video sampling improves motion coherence and physics alignment by using a pre-trained video generator as its own denoising autoencoder for iterative refinement with uncertainty-aware region selection.
Haocheng Xi, Charlie Ruan, Peiyuan Liao +7 authors
Quantized reinforcement learning training using FP8 precision faces stability issues due to numerical mismatches between training and inference phases, but a unified FP8 framework achieves significant speedups with stable convergence.
Kaixuan Fan, Kaituo Feng, Manyuan Zhang +7 authors
Agent-RRM, a multi-faceted reward model, provides structured feedback for agentic trajectories through reasoning traces, critiques, and performance scores, with unified feedback integration showing superior performance across diverse benchmarks.
Xilin Jiang, Qiaolin Wang, Junkai Wu +30 authors
Current multimodal models demonstrate limited understanding of cultural and contextual audio-visual content, particularly excelling only in surface-level analysis rather than deeper semantic comprehension.
Justin Cui, Jie Wu, Ming Li +6 authors
Researchers developed a method to overcome sink-collapse in autoregressive video generation by addressing the conflict between Rotary Position Embedding and multi-head attention mechanisms, enabling real-time streaming of videos up to 12 hours long.
Jonathan Kahana, Eliahu Horwitz, Yedid Hoshen
Hidden superior models exist in public repositories but are overlooked due to inefficient discovery methods; a multi-armed bandit approach using shared query sets and aggressive elimination significantly accelerates identification of top-performing models.
Aniket Rege, Arka Sadhu, Yuliang Li +5 authors
An agentic framework using entity scene graphs enables long-horizon video understanding with structured search, temporal reasoning, and cross-modal capabilities for extended visual and audio interpretation.
Dingkun Liu, Yuheng Chen, Zhu Chen +5 authors
EEG foundation models show mixed performance compared to specialist models, with larger models not always providing better generalization under current data regimes.
Changjian Jiang, Kerui Ren, Xudong Li +7 authors
PLANING presents an efficient streaming reconstruction framework that combines explicit geometric primitives with neural Gaussians to achieve high-quality rendering and accurate geometry simultaneously through decoupled optimization.
Ellen Xiaoqing Tan, Shehzaad Dhuliawala, Jing Xu +4 authors
A reinforcement learning-based pretraining method improves language model safety, factuality, and quality by evaluating generations through a combination of model rollouts, original suffixes, and rewritten suffixes.
Andrew Kyle Lampinen, Yuxuan Li, Eghbal Hosseini +2 authors
Linear representation directions in language models dynamically shift during conversations, affecting how factual information is encoded while preserving generic content, with implications for interpretability and context-adaptive model behavior.
Kanishk Gandhi, Shivam Garg, Noah D. Goodman +1 authors
Endless Terminals introduces an autonomous pipeline for generating procedural terminal tasks that significantly improves agent performance on both synthetic and human-curated benchmarks through scalable reinforcement learning environments.
Yuxuan Wan, Tianqing Fang, Zaitang Li +5 authors
A novel self-evolving framework for Deep Research Agents that enhances performance through iterative verification and rubric-based feedback during inference, achieving significant accuracy improvements without additional training.
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