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。
Mo Yu, Lemao Liu, Junjie Wu +5 authors
A study investigates whether large language models understand physical concepts through a grid-based task, showing that they significantly underperform compared to humans and highlighting the stochastic parrot phenomenon.
50 篇论文 · 按点赞排序
Jonas Geiping, Sean McLeish, Neel Jain +6 authors
A novel language model scales test-time computation by reasoning in latent space, outperforming token-based models on reasoning benchmarks.
Runze Liu, Junqi Gao, Jian Zhao +5 authors
Test-Time Scaling improves Large Language Model performance by optimizing inference computation based on policy models, Process Reward Models, and problem difficulty.
Heejun Lee, Geon Park, Jaduk Suh +1 authors
InfiniteHiP is a novel LLM inference framework that enhances long-context processing with token pruning, RoPE adjustment, and host memory offloading, enabling faster and more memory-efficient operation.
Kiran Kamble, Melisa Russak, Dmytro Mozolevskyi +3 authors
FailSafeQA evaluates the robustness and context-awareness of large language models in financial applications through domain expertise, query completeness, linguistic accuracy, and document relevance challenges.
Shoufa Chen, Chongjian Ge, Yuqi Zhang +19 authors
Goku, a state-of-the-art family of joint image-and-video generation models using rectified flow Transformers, sets new benchmarks in text-to-image and text-to-video tasks.
Daniil Moskovskiy, Nikita Sushko, Sergey Pletenev +2 authors
A pipeline generates multilingual synthetic detoxification data using LLMs, outperforming models trained on human-annotated datasets in few-shot settings.
Fanxu Meng, Zengwei Yao, Muhan Zhang
TransMLA converts GQA-based models to MLA to reduce KV cache size and improve efficiency without sacrificing expressiveness.
OpenAI, Ahmed El-Kishky, Alexander Wei +22 authors
General-purpose reinforcement learning applied to large language models outperforms domain-specific systems in complex coding and reasoning tasks, achieving top results in competitions without hand-crafted strategies.
Xilin Wei, Xiaoran Liu, Yuhang Zang +9 authors
VideoRoPE, a 3D position embedding that preserves spatio-temporal relationships, improves performance on video tasks by addressing limitations in previous RoPE variants.
Lingfei Qian, Weipeng Zhou, Yan Wang +3 authors
A study evaluates 16 large language models on complex financial tasks, finding that domain-specific CoT fine-tuning and reinforcement learning improve performance and highlight the need for further research on long-context and multi-table reasoning.
Chengqi Lyu, Songyang Gao, Yuzhe Gu +14 authors
A novel reinforcement learning framework, OREAL, using outcome rewards achieves high accuracy in mathematical reasoning tasks with smaller models compared to larger models trained using distillation.
Xu Huang, Wenhao Zhu, Hanxu Hu +4 authors
BenchMAX evaluates multilingual large language models across advanced capabilities like instruction following, reasoning, and code generation, revealing language-specific performance gaps.
Junlong Li, Daya Guo, Dejian Yang +3 authors
CodeI/O transforms code into a code input-output prediction format to train models on universal reasoning primitives and improve performance across various reasoning tasks through CoT rationales and multi-turn revision.
Peiyuan Zhang, Yongqi Chen, Runlong Su +4 authors
Sliding tile attention (STA) accelerates 3D attention in Diffusion Transformers (DiTs), reducing video generation latency without quality loss.
Dan Busbridge, Amitis Shidani, Floris Weers +3 authors
The study presents a scaling law to optimize compute allocation for model distillation, showing conditions under which distillation outperforms supervised pretraining.
Alex Jinpeng Wang, Dongxing Mao, Jiawei Zhang +9 authors
TextAtlas5M is a novel dataset of 5 million images for evaluating long-text rendering in text-conditioned image generation, presenting challenges even for advanced proprietary models.
Andrei Panferov, Jiale Chen, Soroush Tabesh +3 authors
QuEST, a new method for quantization-aware training, achieves better accuracy at lower model size with weights and activations in 4-bits or less, even stable training with 1-bit, by improving quantization accuracy and the gradient estimator.
Yujie Zhou, Jiazi Bu, Pengyang Ling +10 authors
Light-A-Video achieves temporally smooth video relighting through a training-free approach, using a Consistent Light Attention module and Progressive Light Fusion strategy to enhance lighting consistency and maintain image quality.
Qinghe Wang, Yawen Luo, Xiaoyu Shi +7 authors
CineMaster is a framework for 3D-aware and controllable text-to-video generation using user-defined control signals and a text-to-video diffusion model.
Hoigi Seo, Wongi Jeong, Jae-sun Seo +1 authors
Skip and Re-use layers (Skrr) enhance memory efficiency in text encoders of T2I diffusion models without affecting image quality.
Dacheng Li, Shiyi Cao, Tyler Griggs +6 authors
Data-efficient supervised fine-tuning and parameter-efficient low-rank adaptation enable large language models to learn long chain-of-thought reasoning effectively, with structural correctness being more crucial than content accuracy in training samples.
Mengxi Xiao, Zihao Jiang, Lingfei Qian +10 authors
A retrieval-augmented generation framework enhances stock prediction accuracy by fine-tuning a large language model and integrating it with a novel retrieval method on specialized datasets.
Wenfang Sun, Xinyuan Song, Pengxiang Li +3 authors
LayerNorm Scaling addresses the Curse of Depth in LLMs by scaling variance inversely with layer depth, improving pre-training and fine-tuning effectiveness of deep layers.
Chung-Ho Wu, Yang-Jung Chen, Ying-Huan Chen +8 authors
AuraFusion360 enhances 3D scene inpainting in 360-degree unbounded scenes using Gaussian Splatting, depth-aware masking, Adaptive Guided Depth Diffusion, and SDEdit for superior quality and geometric accuracy.
Yangguang Li, Zi-Xin Zou, Zexiang Liu +8 authors
TripoSG, a novel shape diffusion framework, generates high-fidelity 3D meshes from images using a rectified flow transformer and hybrid training strategy.
Hongwei Yi, Shitong Shao, Tian Ye +7 authors
Magic141 efficiently generates high-quality videos by factorizing the text-to-video task into text-to-image and image-to-video components, optimizing for memory, speed, and quality.
Bidipta Sarkar, Warren Xia, C. Karen Liu +1 authors
Language models trained without human demonstrations improve multi-agent communication skills through reinforcement learning, enabling effective zero-shot coordination in complex social deduction tasks.
Jonathan Kahana, Or Nathan, Eliahu Horwitz +1 authors
ProbeLog is a method for classification model retrieval based on probe logit analysis, offering efficient and scalable search without metadata.
Yung-Sung Chuang, Benjamin Cohen-Wang, Shannon Zejiang Shen +6 authors
SelfCite is a self-supervised method that uses context ablation to align LLMs for generating high-quality, sentence-level citations, improving citation quality and F1 scores on long-form question answering 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