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。
Yiping Wang, Qing Yang, Zhiyuan Zeng +11 authors
Reinforcement learning with verifiable reward using one training example significantly enhances math reasoning capabilities of large language models.
50 篇论文 · 按点赞排序
Zhiqiu Lin, Siyuan Cen, Daniel Jiang +12 authors
CameraBench assesses and enhances camera motion understanding by providing a large-scale dataset, taxonomy, and benchmark for evaluating SfM and VLMs, emphasizing the need for both semantic and geometric information.
Shivalika Singh, Yiyang Nan, Alex Wang +10 authors
Measuring progress is fundamental to the advancement of any scientific field. As benchmarks play an increasingly central role, they also grow more susceptible to distortion. Chatbot Arena has emerged as the go-to leaderboard for ranking the most capable AI systems. Yet, in this work we identify systematic issues that have resulted in a distorted playing field. We find that undisclosed private testing practices benefit a handful of providers who are able to test multiple variants before public release and retract scores if desired. We establish that the ability of these providers to choose the best score leads to biased Arena scores due to selective disclosure of performance results. At an extreme, we identify 27 private LLM variants tested by Meta in the lead-up to the Llama-4 release. We also establish that proprietary closed models are sampled at higher rates (number of battles) and have fewer models removed from the arena than open-weight and open-source alternatives. Both these policies lead to large data access asymmetries over time. Providers like Google and OpenAI have received an estimated 19.2% and 20.4% of all data on the arena, respectively. In contrast, a combined 83 open-weight models have only received an estimated 29.7% of the total data. We show that access to Chatbot Arena data yields substantial benefits; even limited additional data can result in relative performance gains of up to 112% on the arena distribution, based on our conservative estimates. Together, these dynamics result in overfitting to Arena-specific dynamics rather than general model quality. The Arena builds on the substantial efforts of both the organizers and an open community that maintains this valuable evaluation platform. We offer actionable recommendations to reform the Chatbot Arena's evaluation framework and promote fairer, more transparent benchmarking for the field
Prateek Chhikara, Dev Khant, Saket Aryan +2 authors
Mem0, a memory-centric architecture with graph-based memory, enhances long-term conversational coherence in LLMs by efficiently extracting, consolidating, and retrieving information, outperforming existing memory systems in terms of accuracy and computational efficiency.
Woongyeong Yeo, Kangsan Kim, Soyeong Jeong +2 authors
UniversalRAG enhances Retrieval-Augmented Generation by dynamically selecting and retrieving knowledge from various heterogeneous and modality-specific sources to improve factual accuracy across different types of queries.
Xiaoxi Li, Jiajie Jin, Guanting Dong +5 authors
WebThinker, an autonomous deep research agent, improves LRM performance on complex tasks by integrating a Deep Web Explorer for dynamic web information gathering and using RL-based training through Direct Preference Optimization.
Zeina Aldallal, Sara Chrouf, Khalil Hennara +3 authors
Sadeed, a fine-tuned decoder-only language model, enhances Arabic text diacritization using limited resources and addresses benchmarking limitations with SadeedDiac-25.
Chris, Yichen Wei, Yi Peng +10 authors
Skywork R1V2 enhances multimodal reasoning through a hybrid reinforcement learning approach that balances reward-model guidance and rule-based strategies, improving training efficiency with the Selective Sample Buffer mechanism and mitigating visual hallucinations.
Marah Abdin, Sahaj Agarwal, Ahmed Awadallah +20 authors
Phi-4-reasoning, a 14-billion parameter model enhanced with supervised fine-tuning and reinforcement learning, outperforms larger models on complex reasoning tasks across various benchmarks.
Wenkai Yang, Jingwen Chen, Yankai Lin +1 authors
A novel two-stage framework using Qwen2.5-72B-Instruct enhances LLMs' math critique ability by generating detailed step-wise critiques and applying reinforcement learning, resulting in better error identification and refinement.
Rulin Shao, Rui Qiao, Varsha Kishore +8 authors
ReasonIR-8B, a retriever trained with synthetic challenging queries, achieves state-of-the-art results in reasoning-intensive information retrieval and RAG tasks, improving performance and computational efficiency.
Hongyu Wang, Shuming Ma, Furu Wei
BitNet v2 enables native 4-bit activation quantization for 1-bit Large Language Models by using an online Hadamard transformation to mitigate activation outliers, achieving performance comparable to 8-bit activations with reduced memory and computational costs.
Haoran Xu, Baolin Peng, Hany Awadalla +11 authors
A structured training method with high-quality Chain-of-Thought data improves reasoning abilities in small language models.
Jiwen Yu, Yiran Qin, Haoxuan Che +7 authors
Interactive Generative Video (IGV) combines generative capabilities with interactive features to enable diverse applications across gaming, embodied AI, and autonomous driving, focusing on challenges such as real-time generation, control, memory, dynamics, and intelligence.
Dongzhi Jiang, Ziyu Guo, Renrui Zhang +6 authors
T2I-R1, a reasoning-enhanced text-to-image generator using RL and bi-level chain-of-thought reasoning, improves performance by 13% on T2I-CompBench and 19% on WISE compared to Janus-Pro.
Zae Myung Kim, Chanwoo Park, Vipul Raheja +1 authors
MPO is a framework that dynamically refines reward signals in LLMs through meta-reward modeling, enhancing alignment and reducing reward hacking and prompt engineering needs.
Chong Zhang, Yue Deng, Xiang Lin +8 authors
Recent replication studies explore the supervised fine-tuning and reinforcement learning from verifiable rewards of reasoning language models, focusing on data preparation and method design, and highlight challenges and potential enhancements.
Zayd M. K. Zuhri, Erland Hilman Fuadi, Alham Fikri Aji
Softpick is a drop-in replacement for softmax in transformer attention mechanisms that improves performance, reduces activation sinks, and achieves sparsity, particularly beneficial for quantization and low-precision training.
Haofan Wang, Yujia Xu, Yimeng Li +5 authors
RepText enhances pre-trained monolingual text-to-image models to generate precise multilingual text in specified fonts using ControlNet and a text perceptual loss.
Xindi Wu, Hee Seung Hwang, Polina Kirichenko +1 authors
COMPACT, a compositional visual capability tuning method, improves multimodal large language models' performance on complex vision-language tasks with less data than traditional visual instruction tuning.
Vishnu Sarukkai, Zhiqiang Xie, Kayvon Fatahalian
LLM agents can improve performance on sequential decision-making tasks by learning from self-generated examples, enhancing test performance across multiple benchmarks without task-specific knowledge engineering.
Andrew M. Bean, Rebecca Payne, Guy Parsons +8 authors
LLMs perform well in medical licensing exams but fail to provide accurate medical advice to the public due to challenges in user interaction.
Hasan Abed Al Kader Hammoud, Hani Itani, Bernard Ghanem
Aggregating answers from multiple subthoughts in reasoning traces improves the accuracy of Large Language Models beyond relying on the final answer.
Guangyi Liu, Pengxiang Zhao, Liang Liu +15 authors
LLM-driven phone GUI agents evolve from script-based automation to intelligent systems, using advanced language understanding, multimodal perception, and robust decision-making to address generality, maintenance overhead, and intent comprehension challenges.
Haoyu Zhen, Qiao Sun, Hongxin Zhang +4 authors
The paper introduces a 4D world model trained on RGB-DN videos to predict spatial and temporal dynamics in embodied environments, improving inverse dynamics learning, view synthesis, and policy performance.
KimiTeam, Ding Ding, Zeqian Ju +37 authors
Kimi-Audio, an open-source audio foundation model, achieves state-of-the-art performance across audio-related tasks through a novel LLM-based architecture and comprehensive training and evaluation processes.
Paul Kassianik, Baturay Saglam, Alexander Chen +15 authors
Foundation-Sec-8B, a LLM enhanced with cybersecurity-specific training, matches high-performance models in cybersecurity tasks and promotes AI adoption in the field.
Zechuan Zhang, Ji Xie, Yu Lu +2 authors
A novel in-context editing framework using a Diffusion Transformer enhances instruction-following precision and efficiency without extensive retraining or large datasets.
Mohammad Mahdi Abootorabi, Omid Ghahroodi, Pardis Sadat Zahraei +17 authors
This survey provides a comprehensive overview of generative AI applications in character animation, covering facial animation, expression rendering, image synthesis, avatar creation, gesture modeling, motion synthesis, object generation, and texture synthesis.
Antoni Bigata, Rodrigo Mira, Stella Bounareli +4 authors
KeySync, a two-stage framework, addresses expression leakage and occlusions in lip synchronization, improving visual quality and achieving state-of-the-art results in lip reconstruction and cross-synchronization.
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