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。
Tianzhe Chu, Yuexiang Zhai, Jihan Yang +6 authors
Reinforcement learning demonstrates superior generalization compared to supervised fine-tuning across textual and visual domains, while supervised fine-tuning stabilizes the model output format essential for RL performance.
50 篇论文 · 按点赞排序
Yue Liu, Hongcheng Gao, Shengfang Zhai +7 authors
GuardReasoner enhances LLM safety through enhanced reasoning capabilities using specialized training datasets and techniques, outperforming existing models on guardrail tasks.
Long Phan, Alice Gatti, Ziwen Han +1155 authors
HLE is a challenging multi-modal benchmark that highlights the limitations of current LLMs in closed-ended academic questions.
An Yang, Bowen Yu, Chengyuan Li +25 authors
The Qwen2.5-1M series models extend context length to 1 million tokens with enhanced long-context capabilities, employing techniques like long data synthesis and progressive pre-training, and are supported by an open-source inference framework with sparse attention and kernel optimizations.
Yadong Li, Jun Liu, Tao Zhang +90 authors
Baichuan-Omni-1.5 is an omni-modal model with end-to-end audio generation, featuring a comprehensive data pipeline, audio-tokenizer, and multi-stage training strategy for superior performance across multimodal tasks.
Yue Wang, Qiuzhi Liu, Jiahao Xu +11 authors
A new decoding strategy addresses underthinking in large language models by penalizing frequent thought switching, improving accuracy without fine-tuning.
Yubo Wang, Xiang Yue, Wenhu Chen
Critique Fine-Tuning (CFT) improves language model performance by training them to critique noisy responses, showing better reasoning ability than standard Supervised Fine-Tuning (SFT).
Liang Wang, Haonan Chen, Nan Yang +3 authors
CoRAG, a multi-step retrieval and reasoning approach, enhances RAG models by dynamically refining queries and using rejection sampling to improve performance, especially in multi-hop question answering.
Ruizhe Wang, Yeyun Gong, Xiao Liu +5 authors
The FP4 training framework for LLMs includes differentiable quantization and outlier clamping to achieve accuracy comparable to BF16 and FP8 with minimal degradation on large-scale models.
Andrei Alexandru, Antonia Calvi, Henry Broomfield +9 authors
Atla Selene Mini, an 8B language model-as-a-judge, excels across various benchmarks using enhanced data curation and a combined training approach, achieving top performance in zero-shot evaluations and real-world scenarios.
Hongzhi Huang, Defa Zhu, Banggu Wu +4 authors
Over-Tokenized Transformers improve language modeling by decoupling input and output vocabularies, leading to better performance with larger input vocabularies.
Scott Fujimoto, Pierluca D'Oro, Amy Zhang +2 authors
The paper presents MR.Q, a model-free deep RL algorithm that uses model-based representations to improve performance across diverse benchmarks without incurring high computational costs.
Arthur Douillard, Yanislav Donchev, Keith Rush +11 authors
The paper presents enhancements to DiLoCo for distributed training of large language models, reducing peak bandwidth and wall clock time while maintaining learning quality.
Micah Rentschler, Jesse Roberts
A pre-trained transformer fine-tuned with reinforcement learning develops the ability to solve unseen problems with sample efficiency and adaptability, demonstrating robust meta-learning.
Zicheng Zhang, Xiangyu Zhao, Xinyu Fang +6 authors
With the rapid iteration of Multi-modality Large Language Models (MLLMs) and the evolving demands of the field, the number of benchmarks produced annually has surged into the hundreds. The rapid growth has inevitably led to significant redundancy among benchmarks. Therefore, it is crucial to take a step back and critically assess the current state of redundancy and propose targeted principles for constructing effective MLLM benchmarks. In this paper, we focus on redundancy from three key perspectives: 1) Redundancy of benchmark capability dimensions, 2) Redundancy in the number of test questions, and 3) Cross-benchmark redundancy within specific domains. Through the comprehensive analysis over hundreds of MLLMs' performance across more than 20 benchmarks, we aim to quantitatively measure the level of redundancy lies in existing MLLM evaluations, provide valuable insights to guide the future development of MLLM benchmarks, and offer strategies to refine and address redundancy issues effectively.
Enze Xie, Junsong Chen, Yuyang Zhao +10 authors
SANA-1.5, a linear Diffusion Transformer, achieves state-of-the-art text-to-image generation with efficient scaling through depth growth, model pruning, and inference-time scaling.
Zhengyang Tang, Ziniu Li, Zhenyang Xiao +8 authors
A new benchmark evaluates LLM critique capabilities using closed-loop methodology with self-critique, cross-critique, and iterative critique, revealing differences between classical and advanced reasoning models.
Yuxin Zuo, Shang Qu, Yifei Li +6 authors
MedXpertQA is a benchmark for evaluating medical expertise and reasoning with a focus on advanced clinical knowledge and diverse data, including images and patient records.
Lin Yueyu, Li Zhiyuan, Peter Yue +1 authors
A series of models distilled from Qwen 2.5 using RWKV-7 attention architecture enhance RNN expressiveness and demonstrate improved state tracking compared to transformers, while reducing knowledge processing time and enabling transfer from larger models.
Lan Pan, Hanbo Xie, Robert C. Wilson
LLMs generally underperform humans in exploration tasks, relying more on uncertainty-driven strategies and making premature decisions, unlike humans who balance uncertainty and empowerment.
Aitor Arrieta, Miriam Ugarte, Pablo Valle +2 authors
DeepSeek-R1 and OpenAI's o3-mini are assessed for safety using ASTRAL, with OpenAI's model showing significantly higher safety performance.
Chenguo Lin, Panwang Pan, Bangbang Yang +2 authors
DiffSplat is a 3D generative model that uses text-to-image diffusion models to generate consistent 3D Gaussian splats from text or images, leveraging web-scale 2D priors and introducing a 3D rendering loss for coherence.
Lee Sharkey, Bilal Chughtai, Joshua Batson +26 authors
Mechanistic interpretability aims to understand neural network mechanisms to enhance AI assurance and address scientific questions about intelligence, highlighting open problems in methods, application, and socio-technical challenges.
Benjamin Feuer, Chinmay Hegde
The WILDCHAT-50M dataset, comprising responses from over 50 open-weight models, enables large-scale comparative analysis and demonstrates improved performance in SFT mixtures with fewer samples.
Haorui He, Zengqiang Shang, Chaoren Wang +11 authors
Emilia-Pipe preprocessing pipeline creates Emilia and Emilia-Large datasets, enhancing speech generation with high-quality in-the-wild multilingual data, surpassing audiobook datasets.
Wei Chow, Jiageng Mao, Boyi Li +3 authors
A benchmark and framework, PhysBench and PhysAgent, enhance Vision-Language Models' understanding of physical phenomena in real-world environments.
Hailong Guo, Bohan Zeng, Yiren Song +3 authors
Any2AnyTryon generates high-quality, flexible, and controllable image-based virtual try-ons using adaptive position embedding and a large-scale synthetic dataset.
Jenna Russell, Marzena Karpinska, Mohit Iyyer
In this paper, we study how well humans can detect text generated by commercial LLMs (GPT-4o, Claude, o1). We hire annotators to read 300 non-fiction English articles, label them as either human-written or AI-generated, and provide paragraph-length explanations for their decisions. Our experiments show that annotators who frequently use LLMs for writing tasks excel at detecting AI-generated text, even without any specialized training or feedback. In fact, the majority vote among five such "expert" annotators misclassifies only 1 of 300 articles, significantly outperforming most commercial and open-source detectors we evaluated even in the presence of evasion tactics like paraphrasing and humanization. Qualitative analysis of the experts' free-form explanations shows that while they rely heavily on specific lexical clues ('AI vocabulary'), they also pick up on more complex phenomena within the text (e.g., formality, originality, clarity) that are challenging to assess for automatic detectors. We release our annotated dataset and code to spur future research into both human and automated detection of AI-generated text.
Clément Desroches, Martin Chauvin, Louis Ladan +3 authors
The paper presents a methodology to estimate the environmental impact of a company's AI portfolio, highlighting the significant energy consumption of large generative AI models and advocating for coordinated efforts and standardized frameworks to align AI development with net-zero goals.
Aitor Arrieta, Miriam Ugarte, Pablo Valle +2 authors
Researchers used ASTRAL to generate and test unsafe inputs on OpenAI's o3-mini LLM, identifying 87 instances of unsafe behavior before deployment.
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