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。
Junyu Luo, Xiao Luo, Kaize Ding +3 authors
A noise-robust supervised fine-tuning framework enhances large language models' performance in downstream tasks by detecting and relabeling noisy data.
50 篇论文 · 按点赞排序
Yiwen Hu, Huatong Song, Jia Deng +8 authors
YuLan-Mini, a 2.42B parameter base model, achieves top-tier performance with efficient pre-training techniques, including a data pipeline with cleaning and scheduling, robust optimization, and annealing with targeted data selection.
Yuqing Wang, Shuhuai Ren, Zhijie Lin +6 authors
A parallel generation strategy for autoregressive models improves inference speed without significantly compromising quality in visual generation tasks.
Weihao Zeng, Yuzhen Huang, Lulu Zhao +3 authors
B-STaR framework autonomously adjusts configurations to enhance exploration and exploitation, improving performance in self-improving models.
Tingxu Han, Chunrong Fang, Shiyu Zhao +3 authors
A token-budget-aware framework dynamically estimates and allocates token budgets for LLM reasoning, reducing costs with minimal performance loss.
Wei Liu, Junlong Li, Xiwen Zhang +3 authors
A study on self-evolving training for multimodal reasoning identifies key factors and configurations to optimize reasoning abilities without additional human annotations, introducing MSTaR as an effective training framework.
Ermo Hua, Che Jiang, Xingtai Lv +7 authors
New Fourier Position Embedding (FoPE) improves the frequency-domain properties of attention mechanisms, enhancing length generalization in language models compared to existing methods like Rotary Position Embedding (RoPE).
Huanjin Yao, Jiaxing Huang, Wenhao Wu +8 authors
A new learning-to-reason method called Collective Monte Carlo Tree Search (CoMCTS) enhances the reasoning capabilities of MLLMs by leveraging collective knowledge, demonstrated through experiments on various benchmarks with Mulberry-260k dataset and Mulberry model.
OpenAI, Aaron Jaech, Adam Kalai +262 authors
The o1 model series, trained with reinforcement learning and chain of thought, enhances safety and robustness by reasoning about policies, leading to superior performance on risk benchmarks while highlighting the need for robust alignment and risk management.
Huaijie Wang, Shibo Hao, Hanze Dong +4 authors
OREO, an offline reinforcement learning method, enhances multi-step reasoning in large language models by optimizing a policy model and value function, improving performance on benchmarks without the need for paired preference data.
Enshu Liu, Xuefei Ning, Yu Wang +1 authors
Distilled Decoding accelerates autoregressive model generation by enabling one-step or two-step output, with improved efficiency and acceptable quality, challenging the notion that AR models are necessarily slow.
Tatiana Zemskova, Dmitry Yudin
A 3D scene graph representation integrated with large language models enhances user-robot interaction by incorporating semantic relationships between objects, leading to improved performance in 3D vision-language tasks compared to baselines.
Zhiheng Liu, Ka Leong Cheng, Qiuyu Wang +7 authors
DepthLab, a depth inpainting model utilizing image diffusion priors, effectively handles missing depth data, preserving scale consistency and excelling in tasks like 3D scene inpainting and LiDAR depth completion.
Chenlong Deng, Zhisong Zhang, Kelong Mao +4 authors
Gist-based context compression in large language models offers near-lossless performance on some tasks but fails on synthetic recall, with identified failure patterns and proposed strategies for improvement.
Luyang Liu, Jonas Pfeiffer, Jiaxing Wu +2 authors
A language model is augmented with an offline coprocessor that optimizes its key-value cache, improving reasoning performance and reducing perplexity.
Jinheon Baek, Sun Jae Lee, Prakhar Gupta +4 authors
Long Context Language Models reduce the sensitivity of In-Context Learning to example selection techniques and instead emphasize the importance of sufficient and augmented examples.
Yazhou Xing, Yang Fei, Yingqing He +4 authors
A novel video Variational Autoencoder (VAE) integrates temporal-aware spatial compression and motion compression with text guidance to achieve high-fidelity video encoding and superior reconstruction quality.
LearnLM Team, Abhinit Modi, Aditya Srikanth Veerubhotla +43 authors
Pedagogical instruction following enhances generative AI systems for educational purposes, leading to a preferred model (LearnLM) across various learning scenarios.
Songhua Liu, Zhenxiong Tan, Xinchao Wang
A convolution-like local attention strategy named CLEAR is introduced to reduce the quadratic complexity of Diffusion Transformers (DiT) to linear, significantly accelerating high-resolution image generation.
Jiaan Wang, Fandong Meng, Yunlong Liang +1 authors
DRT-o1, a neural machine translation model using a multi-agent framework, improves translation quality by simulating long chain-of-thought reasoning, particularly in handling texts with similes and metaphors.
Jialong Wu, Zhenglin Wang, Linhai Zhang +3 authors
SCOPE is a framework that optimizes KV cache separately during prefill and decoding phases to improve long-context and long-output generation in LLMs.
Ho Kei Cheng, Masato Ishii, Akio Hayakawa +3 authors
A multimodal framework jointly trains with video and text to generate high-quality, semantically aligned, and visually synchronized audio, achieving state-of-the-art performance.
Minghao Chen, Roman Shapovalov, Iro Laina +4 authors
PartGen generates 3D objects composed of meaningful parts from text, images, or unstructured 3D assets using multi-view diffusion models for segmentation and completion.
Minghong Cai, Xiaodong Cun, Xiaoyu Li +5 authors
DiTCtrl, a training-free method for multi-prompt video generation, enhances MM-DiT architectures by enabling smooth transitions and consistent motion using an analyzed attention mechanism.
Zhuohao Yu, Weizheng Gu, Yidong Wang +4 authors
Outcome-Refining Process Supervision uses concrete execution signals and tree-structured exploration to enhance algorithmic reasoning in code generation tasks.
Wan-Cyuan Fan, Tanzila Rahman, Leonid Sigal
MMFactory is a universal framework that combines various models and metrics to propose tailored solutions for visual tasks based on user specifications and constraints.
Sungjin Park, Xiao Liu, Yeyun Gong +1 authors
A new ensemble framework combining language models with Monte Carlo Tree Search improves performance on complex mathematical reasoning tasks.
Jinhui Yi, Syed Talal Wasim, Yanan Luo +2 authors
An encoder-free approach for video-language understanding using Spatio-Temporal Alignment Blocks achieves competitive performance with significantly reduced computational overhead and faster processing speeds.
Łukasz Borchmann
The evaluation setup in ARC Challenge affects perceived model difficulty, leading to false implications of reasoning deficits, whereas fairer methods reduce performance gaps and reveal superhuman capabilities in tasks like OpenBookQA.
Yucong Luo, Qitao Qin, Hao Zhang +4 authors
Molar, a multimodal large language sequential recommendation framework, enhances recommendation performance by integrating textual and non-textual data with ID information to capture both collaborative filters and user interests.
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