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
Elliot Meyerson, Giuseppe Paolo, Roberto Dailey +6 authors
MAKER, a system using microagents with error correction, successfully solves tasks with over a million LLM steps, suggesting a new approach for scaling LLM capabilities.
Fangqi Zhu, Zhengyang Yan, Zicong Hong +3 authors
WMPO, a pixel-based world-model framework for VLA RL, enhances sample efficiency, performance, self-correction, and generalization in robotic manipulation tasks.
Sean McLeish, Ang Li, John Kirchenbauer +7 authors
Converting pretrained non-recurrent language models to depth-recurrent models improves performance at a given compute budget using a curriculum of recurrences.
Sreyan Ghosh, Arushi Goel, Lasha Koroshinadze +8 authors
Music Flamingo, a large audio-language model, advances music understanding through fine-tuning on a rich dataset and post-training with novel methods, achieving state-of-the-art results across various benchmarks.
Fei Zhao, Chonggang Lu, Haofu Qian +9 authors
RedOne 2.0, a social networking service-oriented LLM, uses a progressive, RL-prioritized post-training paradigm to achieve rapid and stable adaptation, delivering improvements over larger baselines with less data.
Jingwei Ni, Ekaterina Fadeeva, Tianyi Wu +8 authors
Transformer-based uncertainty quantification heads improve step-level reasoning verification in LLMs by estimating uncertainty from internal states, offering a lightweight and scalable alternative to existing methods.
Zhaojian Yu, Kaiyue Feng, Yilun Zhao +3 authors
AlphaResearch, an autonomous research agent, discovers new algorithms in open-ended problems with a dual research environment, achieving competitive performance against human researchers in a benchmark competition.
Kangning Zhang, Wenxiang Jiao, Kounianhua Du +5 authors
A fully automated data evolution framework, LoopTool, enhances tool-use capabilities of Large Language Models by iteratively refining data and model through a closed-loop process.
Zhiyuan Zeng, Hamish Ivison, Yiping Wang +14 authors
Reinforcement Learning with Adaptive Verifiable Environments (RLVE) improves language model reasoning by dynamically adjusting problem difficulty, outperforming static environments and traditional RL training.
Tianhao Peng, Haochen Wang, Yuanxing Zhang +13 authors
MVU-Eval is a comprehensive benchmark for evaluating multi-video understanding in Multimodal Large Language Models, addressing gaps in existing single-video benchmarks and highlighting performance discrepancies in real-world applications.
Zhi Zheng, Wee Sun Lee
A novel policy optimization algorithm, SofT-GRPO, enhances soft-thinking in Large Language Models by integrating Gumbel noise and the Gumbel-Softmax technique, leading to improved performance over discrete-token methods.
Jon Saad-Falcon, Avanika Narayan, Hakki Orhun Akengin +12 authors
Local inference using small language models on accelerators can accurately handle many real-world queries, significantly reducing demand on centralized cloud infrastructure, as measured by intelligence per watt.
Yauhen Babakhin, Radek Osmulski, Ronay Ak +5 authors
A fully open-source text embedding model achieves state-of-the-art performance across embedding tasks, particularly in multilingual scenarios, through a novel data mix and detailed ablation studies.
Zhicheng Cai, Xinyuan Guo, Yu Pei +6 authors
FLEX, a gradient-free learning paradigm, enables Large Language Model agents to continuously evolve through experience, improving performance in tasks like mathematical reasoning, chemical retrosynthesis, and protein fitness prediction.
Mingde Xu, Zhen Yang, Wenyi Hong +6 authors
WebVIA is an agentic framework for generating executable and interactive UI code from design mockups, improving stability and accuracy over existing models.
Muhammad Usama, Mohammad Sadil Khan, Didier Stricker +1 authors
NURBGen generates high-fidelity 3D CAD models from text using Non-Uniform Rational B-Splines, outperforming existing methods in geometric fidelity and dimensional accuracy.
Jinhao Chen, Zhen Yang, Jianxin Shi +2 authors
A proposed Mathematical Self-Evolving framework iteratively refines multimodal large language models through inference, reflection, and reward-based feedback, achieving superior performance in mathematical reasoning tasks compared to existing models.
Yule Wen, Yixin Ye, Yanzhe Zhang +2 authors
Real-Time Reasoning Gym demonstrates the challenges of deploying language models in dynamic environments, introducing AgileThinker to balance reasoning depth and response latency.
Arman Zarei, Samyadeep Basu, Mobina Pournemat +3 authors
SliderEdit enables continuous, fine-grained control over image editing instructions by using low-rank adaptation matrices, improving edit controllability, visual consistency, and user steerability.
Shiyao Xu, Benedetta Liberatori, Gül Varol +1 authors
A new task, Dense Motion Captioning, is introduced with a large-scale dataset, CompMo, and a model, DEMO, that integrates a language model with a motion adapter to generate detailed, temporally grounded captions for 3D human motion sequences.
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