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。
Xuezhi Wang, Denny Zhou
LLMs can perform chain-of-thought reasoning through top-k decoding without manual prompt engineering, outperforming greedy decoding and showing higher confidence in answers.
50 篇论文 · 按点赞排序
Mateusz Łajszczak, Guillermo Cámbara, Yang Li +16 authors
BASE TTS, a large-scale text-to-speech model, demonstrates advanced naturalness and emergent abilities using autoregressive Transformers and novel speech tokenization techniques.
Niklas Muennighoff, Hongjin Su, Liang Wang +5 authors
GrIT allows large language models to excel at both generative and embedding tasks through task instruction, leading to new state-of-the-art performance without sacrificing efficiency.
Shivalika Singh, Freddie Vargus, Daniel Dsouza +30 authors
The initiative builds a human-curated instruction-following dataset spanning 65 languages and creates the largest multilingual collection of instruction-following instances through templating and translating existing datasets across 114 languages, contributing datasets and platforms for participatory research.
Ahmet Üstün, Viraat Aryabumi, Zheng-Xin Yong +14 authors
Aya, a multilingual generative language model supporting over 50% lower-resourced languages, excels in both generative and discriminative tasks across 99 languages, and provides extensive evaluations and open-source resources.
Zhiyong Wu, Chengcheng Han, Zichen Ding +5 authors
OS-Copilot framework enables FRIDAY, a self-improving generalist agent, to automate a wide range of computer tasks with enhanced generalization and minimal supervision.
Noveen Sachdeva, Benjamin Coleman, Wang-Cheng Kang +6 authors
Data-efficient methods like Ask-LLM and Density sampling improve model quality and training efficiency in large language models by optimizing data selection and coverage.
Hao Liu, Wilson Yan, Matei Zaharia +1 authors
Large-scale training of a transformer model for long video and language sequences using RingAttention and masked sequence packing improves performance on retrieval and long-form understanding tasks.
Kuang-Huei Lee, Xinyun Chen, Hiroki Furuta +2 authors
ReadAgent, an LLM-based prompting system, enhances effective context length up to 20x through memory episodes and gist memories, outperforming baselines on long-document reading comprehension tasks.
Shubham Toshniwal, Ivan Moshkov, Sean Narenthiran +3 authors
A new math instruction tuning dataset, OpenMathInstruct-1, leverages the Mixtral model to achieve competitive performance on math reasoning benchmarks using open-source LLMs.
Johan Obando-Ceron, Ghada Sokar, Timon Willi +6 authors
Incorporating Soft MoE modules into value-based networks enhances scalability and performance across different training regimes and model sizes in reinforcement learning.
Huizhuo Yuan, Zixiang Chen, Kaixuan Ji +1 authors
Self-play fine-tuning improves diffusion models' performance and alignment with human preferences using competition between model versions, outperforming supervised and RL-based methods with less data.
Di Zhang, Wei Liu, Qian Tan +12 authors
A dialogue-based large language model for chemistry, ChemLLM, demonstrates superior performance over existing models across various chemical tasks and shows adaptability to mathematical and physical tasks by transforming structured chemical knowledge into conversational format.
Ze Ma, Daquan Zhou, Chun-Hsiao Yeh +6 authors
A framework named Video Custom Diffusion (VCD) is proposed for subject identity controllable video generation, achieving stable and high-quality outputs with better identity preservation compared to existing methods.
Xinyun Chen, Ryan A. Chi, Xuezhi Wang +1 authors
LLMs exhibit reduced performance on reasoning tasks when premises are reordered, affecting both deductive reasoning and mathematical problem-solving.
Ashish Shenoy, Yichao Lu, Srihari Jayakumar +11 authors
Lumos is an end-to-end multimodal question-answering system that integrates Scene Text Recognition and a Multimodal Large Language Model to improve text understanding and efficiency from first-person images.
Yao Fu, Rameswar Panda, Xinyao Niu +4 authors
Continual pretraining with domain-balanced and length-upsampled data effectively scales language models' context length to 128K without significant additional computational cost.
James Liu, Guangxuan Xiao, Kai Li +4 authors
BitDelta quantizes the fine-tuning delta of large language models to 1 bit, reducing GPU memory usage by over 10x without significant performance loss.
Keisuke Kamahori, Yile Gu, Kan Zhu +1 authors
Fiddler is an efficient inference engine for large MoE models that minimizes data movement between CPU and GPU, enabling high-speed inference on resource-constrained settings.
Hila Manor, Tomer Michaeli
Two zero-shot techniques using DDPM inversion on pre-trained diffusion models are explored for editing audio signals, enabling text-based manipulation and discovery of semantically meaningful editing directions.
Huaiyuan Ying, Shuo Zhang, Linyang Li +19 authors
InternLM-Math, a math reasoning LLM, achieves state-of-the-art performance in various benchmarks by integrating chain-of-thought reasoning, reward modeling, formal reasoning, data augmentation, and code interpretation.
Yutaro Yamada, Khyathi Chandu, Yuchen Lin +3 authors
A language agent with chain-of-3D-thoughts (L3GO) enhances 3D object generation for unconventional descriptions by using large language models to compose objects in a 3D simulation environment, outperforming GPT-4 and other models in the UFO benchmark.
Yifan Zhang, Yifan Luo, Yang Yuan +1 authors
A novel strategy using meta-prompted language models autonomously selects high-quality mathematical data for continual pretraining, significantly improving a language model's mathematical reasoning and reducing token usage.
Chaoyun Zhang, Liqun Li, Shilin He +9 authors
UFO is a UI-focused agent using GPT-Vision to automate tasks across Windows applications with natural language commands.
Abdullah Hamdi, Luke Melas-Kyriazi, Guocheng Qian +5 authors
GES, a new representation using Generalized Exponential Function, outperforms Gaussian Splatting in 3D reconstruction and generation by requiring fewer particles and improving memory efficiency and rendering speed.
Ali Behrouz, Farnoosh Hashemi
Graph Mamba Networks, a new framework based on selective State Space Models, achieve high performance in graph representation learning with lower computational cost compared to existing Graph Transformers and Message-Passing Neural Networks.
Soroush Nasiriany, Fei Xia, Wenhao Yu +20 authors
A method called PIVOT enables VLMs to perform iterative visual question answering to generate continuous outputs for tasks like robotic control, enabling zero-shot control without task-specific data.
Naoyuki Kanda, Xiaofei Wang, Sefik Emre Eskimez +12 authors
ELaTE is a zero-shot TTS that generates natural and controllable laughter in speech using a combination of conditional flow-matching and laughter detection.
Elvis Dohmatob, Yunzhen Feng, Pu Yang +2 authors
Theoretical analysis and experiments explore how the integration of synthetic data into training datasets affects neural scaling laws and model performance.
Jisu Nam, Heesu Kim, DongJae Lee +3 authors
DreamMatcher personalizes text-to-image (T2I) models by semantically matching reference concepts to target prompts, improving accuracy and diversity without altering the model's structure.
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