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。
Chenfei Wu, Jiahao Li, Jingren Zhou +36 authors
Qwen-Image, an image generation model, advances text rendering and image editing through a comprehensive data pipeline, progressive training, and dual-encoding mechanism.
50 篇论文 · 按点赞排序
Chengshuai Zhao, Zhen Tan, Pingchuan Ma +5 authors
CoT reasoning in LLMs is found to be limited by the distribution discrepancy between training and test data, suggesting it is not a robust form of reasoning.
Yongliang Wu, Yizhou Zhou, Zhou Ziheng +7 authors
Dynamic Fine-Tuning (DFT) improves the generalization of Large Language Models (LLMs) by dynamically rescaling gradients, outperforming standard Supervised Fine-Tuning (SFT) and showing competitive results in offline reinforcement learning.
Shunyu Liu, Minghao Liu, Huichi Zhou +29 authors
VeriGUI is a novel dataset for evaluating GUI agents in long-horizon tasks, emphasizing long-chain complexity and subtask-level verifiability.
Xufang Luo, Yuge Zhang, Zhiyuan He +5 authors
Agent Lightning is a flexible RL framework for training LLMs in various agents, using a hierarchical RL algorithm and decoupling execution from training to handle complex interactions.
Yuxuan Song, Zheng Zhang, Cheng Luo +19 authors
Seed Diffusion Preview, a discrete-state diffusion language model, achieves fast inference speeds through parallel generation, outperforming Mercury and Gemini Diffusion in speed and quality.
Chengsong Huang, Wenhao Yu, Xiaoyang Wang +6 authors
R-Zero is a self-evolving framework that autonomously generates and learns from its own training data, improving reasoning capabilities in LLMs without human-curated tasks.
Tianqing Fang, Zhisong Zhang, Xiaoyang Wang +10 authors
Cognitive Kernel-Pro is an open-source multi-module agent framework that enhances AI agent robustness and performance through data curation and novel test-time strategies, achieving state-of-the-art results.
Ningning Wang, Xavier Hu, Pai Liu +11 authors
A study on the efficiency-effectiveness trade-off in LLM-driven agent systems identifies optimal agent framework design to reduce costs while maintaining performance.
Yue Liao, Pengfei Zhou, Siyuan Huang +11 authors
Genie Envisioner integrates policy learning, evaluation, and simulation using a video diffusion model and neural simulator for instruction-driven robotic manipulation.
Xinrun Xu, Pi Bu, Ye Wang +7 authors
DeepPHY evaluates Vision Language Models' physical reasoning and control through simulated environments with varying difficulty levels.
Peiyu Wang, Yi Peng, Yimeng Gan +12 authors
Skywork UniPic, a 1.5 billion-parameter autoregressive model, unifies image understanding, text-to-image generation, and image editing with state-of-the-art performance on commodity hardware.
Jinsong Li, Xiaoyi Dong, Yuhang Zang +3 authors
DAEDAL, a novel training-free denoising strategy, enables dynamic length adaptation in Diffusion Large Language Models, improving performance and computational efficiency.
Junjie Wu, Jiangnan Li, Yuqing Li +6 authors
A new training paradigm and situated embedding models (SitEmb) enhance retrieval performance by conditioning short text chunks on broader context windows, outperforming state-of-the-art models with fewer parameters.
Zeyi Sun, Ziyu Liu, Yuhang Zang +5 authors
SEAgent, an agentic self-evolving framework, enables computer-use agents to autonomously master novel software through experiential learning and a curriculum of tasks, achieving superior performance compared to existing methods.
Jianxiong Gao, Zhaoxi Chen, Xian Liu +5 authors
LongVie, an end-to-end autoregressive framework, addresses temporal consistency and visual degradation in ultra-long video generation through unified noise initialization, global control signal normalization, multi-modal control, and degradation-aware training.
Alexander Golubev, Maria Trofimova, Sergei Polezhaev +9 authors
Research on applications of Reinforcement Learning (RL) to Large Language Models (LLMs) has mostly been focused on single-turn problems, such as mathematical reasoning or single-shot code generation. While these problems can be viewed as token-level multi-turn MDPs, this view corresponds to a degenerate case of multi-turn interaction where the environment provides no feedback. This contrasts with many real-world domains, such as software engineering (SWE), which require rich multi-turn interactions with a stateful environment that responds to each action with a non-trivial observation. To bridge this gap, we demonstrate the successful application of RL to this general regime. Using a modified Decoupled Advantage Policy Optimization (DAPO) algorithm, we train an agent based on Qwen2.5-72B-Instruct to solve real-world software engineering tasks. Our approach increases the agent's success rate on the SWE-bench Verified benchmark from a 20% rejection fine-tuned baseline to 39%, without relying on any teacher models. On SWE-rebench, our agent matches or outperforms leading open-weight models such as DeepSeek-V3-0324 and Qwen3-235B-A22B using an identical scaffolding, offering a viable path toward building more capable autonomous agents for complex real-world problems based on open models.
Shuai Wang, Ziteng Gao, Chenhui Zhu +2 authors
Pixel Neural Field Diffusion (PixNerd) achieves high-quality image generation in a single-scale, single-stage process without VAEs or complex pipelines, and extends to text-to-image applications with competitive performance.
Xiangru Tang, Zhuoyun Yu, Jiapeng Chen +12 authors
CellForge, an agentic system using a multi-agent framework, transforms raw single-cell multi-omics data into optimized computational models for virtual cells, outperforming state-of-the-art methods in single-cell perturbation prediction.
Shudong Liu, Hongwei Liu, Junnan Liu +8 authors
CompassVerifier is a lightweight, robust model for verifying LLM outputs across various domains, supported by VerifierBench, a comprehensive benchmark dataset.
Qingyu Ren, Qianyu He, Bowei Zhang +6 authors
A self-supervised RL framework enhances instruction following in reasoning models without external supervision, maintaining reasoning performance and offering scalability and cost-effectiveness.
Sajana Weerawardhena, Paul Kassianik, Blaine Nelson +14 authors
Foundation-Sec-8B-Instruct is a cybersecurity-focused LLM designed for chat-style interactions and instruction-following, outperforming other models in cybersecurity tasks while matching their instruction-following capabilities.
George Bredis, Stanislav Dereka, Viacheslav Sinii +2 authors
A lightweight, hyperparameter-free RL algorithm, VL-DAC, enables VLMs to learn generalized policies from inexpensive simulators, improving performance on real-world benchmarks without sacrificing image understanding accuracy.
Yuhan Zhang, Long Zhuo, Ziyang Chu +5 authors
Hi3DEval is a hierarchical evaluation framework for 3D generative content that combines object-level and part-level assessments, including material realism, using a large-scale dataset and hybrid 3D representations.
Bohan Jiang, Dawei Li, Zhen Tan +2 authors
LLMs can be fine-tuned to generate high-quality, audience-tailored explanations of well-being concepts using Supervised Fine-Tuning and Direct Preference Optimization.
Wenxuan Shen, Mingjia Wang, Yaochen Wang +4 authors
Double-Bench is a large-scale, multilingual, and multimodal evaluation system for document Retrieval-Augmented Generation (RAG) systems, addressing limitations in current benchmarks and providing comprehensive assessments of system components.
Zhongyue Zhang, Jiahua Rao, Jie Zhong +22 authors
AuroBind is a scalable virtual screening framework that fine-tunes atomic-level structural models to predict ligand-bound structures and binding fitness, achieving high hit rates in prospective screens across disease-relevant targets.
Jueon Park, Yein Park, Minju Song +4 authors
CoTox, a framework integrating LLMs with chain-of-thought reasoning, enhances multi-toxicity prediction by incorporating chemical structure data, biological pathways, and gene ontology terms, improving interpretability and predictive performance in drug development.
Yi Gui, Zhen Li, Zhongyi Zhang +10 authors
LaTCoder enhances layout preservation in design-to-code tasks by dividing webpage designs into blocks and using Chain-of-Thought reasoning with MLLMs, achieving significant improvements in metrics and human preference.
Haofei Yu, Zhengyang Qi, Yining Zhao +6 authors
Sotopia-RL, a novel reinforcement learning framework, enhances social intelligence in large language models by refining feedback into utterance-level, multi-dimensional rewards, improving performance in social tasks.
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