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。
598 篇论文 · 按点赞排序
Haoran Xu, Amr Sharaf, Yunmo Chen +5 authors
Contrastive Preference Optimization (CPO) enhances moderate-sized LLMs in machine translation, enabling them to match or exceed state-of-the-art model performance.
Keming Lu, Bowen Yu, Chang Zhou +1 authors
Ditto, a self-alignment method, enhances the role-playing capabilities of open-source large language models by fine-tuning them on a large generated dataset, achieving performance comparable to proprietary chatbots.
Zihan Liu, Wei Ping, Rajarshi Roy +3 authors
ChatQA uses a two-stage instruction tuning method with a dense retriever to improve zero-shot conversational QA performance, outperforming GPT-4 on multiple datasets.
Dongsheng Wang, Zhiqiang Ma, Armineh Nourbakhsh +2 authors
A new framework, DocGraphLM, combines pre-trained language models with graph semantics to improve information extraction and question answering on documents with complex layouts.
Ziheng Jiang, Haibin Lin, Yinmin Zhong +29 authors
MegaScale is a production system designed to train large language models at scale using more than 10,000 GPUs, focusing on efficiency, stability, and fault tolerance through advanced observability and diagnostic tools.
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.
Zehan Qi, Xiao Liu, Iat Long Iong +10 authors
WebRL enhances open LLMs as web agents through a self-evolving curriculum, robust outcome-supervised reward model, and adaptive reinforcement learning strategies, achieving high success rates compared to proprietary models.
Enric Corona, Andrei Zanfir, Eduard Gabriel Bazavan +3 authors
VLOGGER generates audio-driven human videos from a single image using a diffusion-based method that includes 3D motion and text-to-image models, outperforming existing methods in quality, identity, and consistency.
Sunny Sanyal, Sujay Sanghavi, Alexandros G. Dimakis
Inheritune leverages transformer blocks from large language models and minimal data to create smaller, efficient base models with performance competitive to larger models.
Chiyu Zhang, Yifei Sun, Jun Chen +7 authors
The SPAR framework enhances content recommendations by using pretrained language models, poly-attention layers, and large language models to effectively process long user engagement histories and predict user-item interactions.
Akari Asai, Jacqueline He, Rulin Shao +22 authors
OpenScholar, a specialized language model, enhances scientific query synthesis by retrieving relevant passages and providing citation-backed answers, outperforming existing models in correctness and citation accuracy.
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.
Dongzhi Jiang, Guanglu Song, Xiaoshi Wu +5 authors
CoMat, an end-to-end diffusion model fine-tuning strategy with an image-to-text concept matching mechanism, improves text-to-image alignment in SDXL without additional data.
Jiongnan Liu, Yutao Zhu, Shuting Wang +6 authors
A novel model constructs user-specific embeddings using a lightweight plug-in module to personalize LLM outputs without fine-tuning, improving performance on various tasks.
Mu Cai, Jianwei Yang, Jianfeng Gao +1 authors
Large Multimodal Models (LMMs) such as LLaVA have shown strong performance in visual-linguistic reasoning. These models first embed images into a fixed large number of visual tokens and then feed them into a Large Language Model (LLM). However, this design causes an excessive number of tokens for dense visual scenarios such as high-resolution images and videos, leading to great inefficiency. While token pruning/merging methods do exist, they produce a single length output for each image and do not afford flexibility in trading off information density v.s. efficiency. Inspired by the concept of Matryoshka Dolls, we propose M3: Matryoshka Multimodal Models, which learns to represent visual content as nested sets of visual tokens that capture information across multiple coarse-to-fine granularities. Our approach offers several unique benefits for LMMs: (1) One can explicitly control the visual granularity per test instance during inference, e.g. , adjusting the number of tokens used to represent an image based on the anticipated complexity or simplicity of the content; (2) M3 provides a framework for analyzing the granularity needed for existing datasets, where we find that COCO-style benchmarks only need around ~9 visual tokens to obtain accuracy similar to that of using all 576 tokens; (3) Our approach provides a foundation to explore the best trade-off between performance and visual token length at sample level, where our investigation reveals that a large gap exists between the oracle upper bound and current fixed-scale representations.
Bishwaranjan Bhattacharjee, Aashka Trivedi, Masayasu Muraoka +31 authors
INDUS, a suite of domain-specific LLMs trained on curated scientific corpora, surpasses general-purpose and existing domain-specific models on specialized NLP tasks and new scientific benchmarks.
Zheng Cai, Maosong Cao, Haojiong Chen +97 authors
InternLM2 is an open-source LLM that outperforms predecessors through innovative pre-training and optimization techniques, including Supervised Fine-Tuning and Conditional Online Reinforcement Learning from Human Feedback.
Han Zhao, Min Zhang, Wei Zhao +3 authors
Cobra, a linear-complexity multimodal large language model, integrates the Mamba language model with visual modality, achieving competitive performance and faster inference speed compared to existing state-of-the-art models.
Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace +8 authors
A large-scale dataset of video-text pairs, automatically curated and annotated using multimodal models, significantly improves performance in video captioning, retrieval, and generation tasks.
Yejie Wang, Keqing He, Dayuan Fu +11 authors
XCoder, a family of models fine-tuned from LLaMA3, achieves state-of-the-art performance on code instruction tasks using a novel data pruning strategy that addresses data leakage and quality in training datasets.
Chunjing Gan, Dan Yang, Binbin Hu +9 authors
MetRag enhances retrieval-augmented generation by integrating similarity and utility-oriented thoughts, employing an LLM as a task adaptive summarizer, and leveraging multi-layered thoughts for knowledge-intensive tasks.
Xueyan Niu, Bo Bai, Lei Deng +1 authors
A theoretical framework using Hopfield networks explains memorization and performance dynamics in Transformer-based language models, showing bounds on achievable cross-entropy loss.
Peitian Zhang, Ninglu Shao, Zheng Liu +4 authors
Llama-3-8B-Instruct's context length is extended from 8K to 80K using QLoRA fine-tuning with minimal additional training samples, demonstrating significant potential for further context extension with increased computational resources.
Zhen Zeng, William Watson, Nicole Cho +4 authors
FlowMind uses Large Language Models with a generic prompt recipe to generate automatic workflows, addressing spontaneous tasks and ensuring data integrity, and it is evaluated using a new financial dataset NCEN-QA.
Raehyuk Jung, Hyojun Go, Jaehyuk Yi +41 authors
Pegasus-1 is a multimodal language model designed for video content comprehension, handling spatiotemporal information and demonstrated in benchmarks for video conversation, zero-shot video question answering, and video summarization.
Luchang Li, Sheng Qian, Jie Lu +3 authors
Transformer-Lite optimizes LLM deployment on device GPUs through various techniques to achieve significant speedups in prefill and decoding over CPU and GPU-based methods.
Chenming Zhu, Tai Wang, Wenwei Zhang +2 authors
LLaVA-3D enhances 2D models with 3D scene understanding by integrating 3D Patch representations and joint 2D/3D instruction tuning, achieving fast convergence and superior performance across tasks.
Viraat Aryabumi, John Dang, Dwarak Talupuru +14 authors
Aya 23 is a powerful multilingual language model for 23 languages, outperforming previous models on a wide range of tasks.
Quentin Garrido, Mahmoud Assran, Nicolas Ballas +3 authors
Image World Models extend Joint-Embedding Predictive Architecture to predict global photometric transformations, offering adaptable and high-performance representations through fine-tuning.
Hongliang He, Wenlin Yao, Kaixin Ma +5 authors
WebVoyager, a Large Multimodal Model, enhances web agent performance by handling multiple modalities and real-world web interactions, achieving high success rates in practical 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