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。
Guilherme Penedo, Hynek Kydlíček, Loubna Ben allal +5 authors
FineWeb, a 15-trillion token dataset from Common Crawl snapshots, outperforms other open datasets in LLM training, and its educational subset, FineWeb-Edu, significantly improves performance on knowledge and reasoning benchmarks.
50 篇论文 · 按点赞排序
Yushun Zhang, Congliang Chen, Ziniu Li +5 authors
Adam-mini reduces memory usage by optimizing learning rate allocation in parameter blocks, achieving comparable or better performance than AdamW with increased throughput.
Ziyan Jiang, Xueguang Ma, Wenhu Chen
The LongRAG framework improves information retrieval and answer extraction by using longer retrieval units and leveraging long-context LLMs, leading to better performance on benchmark datasets.
Shengbang Tong, Ellis Brown, Penghao Wu +11 authors
Cambrian-1, a vision-centric multimodal LLM family, introduces the Spatial Vision Aggregator (SVA) and a new benchmark CV-Bench to enhance visual grounding and representation learning in multimodal systems.
Yuang Peng, Yuxin Cui, Haomiao Tang +7 authors
dreambench++ uses advanced multimodal GPT models to create human-aligned evaluations for generative models in image generation, improving assessment accuracy and efficiency.
Tao Zhang, Xiangtai Li, Hao Fei +5 authors
OMG-LLaVA integrates pixel-level vision understanding and reasoning capabilities through a universal segmentation method combined with LLM, achieving high performance across multiple benchmarks.
Wei Chen, Zhiyuan Li, Zhen Guo +1 authors
An on-device Planner-Action framework, utilizing Phi-3 Mini for planning and Octopus for execution, employs model fine-tuning and multi-LoRA training to enhance efficiency and multi-domain handling on resource-constrained devices.
Terry Yue Zhuo, Minh Chien Vu, Jenny Chim +30 authors
Benchmark Bench assesses LLMs' capability to solve practical programming tasks by invoking function calls from various libraries, revealing significant performance gaps compared to human developers.
Sandeep Mishra, Oindrila Saha, Alan C. Bovik
YouDream generates high-quality, anatomically consistent 3D animals using a text-to-image diffusion model guided by 2D pose priors, showcasing superior creativity and quality compared to prior methods.
Xin Lai, Zhuotao Tian, Yukang Chen +3 authors
A novel method called Step-DPO enhances LLMs' mathematical reasoning by optimizing individual reasoning steps, achieving significant accuracy improvements with minimal data and training.
Aman Singh Thakur, Kartik Choudhary, Venkat Srinik Ramayapally +2 authors
LLM-as-a-judge method for evaluating LLMs shows promise but varies in performance and alignment with human judges, with smaller models like JudgeLM-7B sometimes outperforming larger ones in ranking tasks.
Royi Rassin, Yaron Fairstein, Oren Kalinsky +4 authors
Evaluating retrieval models on partially-annotated datasets can lead to misleading rankings, and more comprehensive annotation improves evaluation reliability.
William Berman, Alexander Peysakhovich
MUMU, a multimodal model combining vision-language encoding and diffusion decoding, synthesizes coherent images from text-image prompts and generalizes to tasks like style transfer and character consistency.
Peiyuan Zhang, Kaichen Zhang, Bo Li +7 authors
By extending the context length of the language backbone, LMMs can process extremely long video sequences, achieving state-of-the-art performance without additional training or complexity.
Zheyuan Zhang, Daniel Zhang-Li, Jifan Yu +5 authors
SimClass, a multi-agent framework with LLM-empowered agents, effectively simulates classroom interactions, enhancing user experience and learning processes through collaboration.
Zijun Yao, Weijian Qi, Liangming Pan +5 authors
SeaKR, an adaptive RAG model, improves knowledge retrieval by leveraging LLMs' self-aware uncertainty to activate, re-rank, and integrate knowledge appropriately, outperforming existing methods in QA tasks.
Zirui Wang, Mengzhou Xia, Luxi He +10 authors
CharXiv evaluates the chart understanding capabilities of MLLMs using a diverse and challenging dataset, uncovering significant weaknesses in model performance compared to human capabilities.
Wenyu Du, Shuang Cheng, Tongxu Luo +5 authors
MIGU, a rehearsal-free and task-label-free method, enhances continual learning in language models by updating parameters with large magnitudes in linear layers, leading to state-of-the-art performance on multiple benchmarks.
Zhenxiong Tan, Xingyi Yang, Songhua Liu +1 authors
Video-Infinity uses distributed inference with Clip parallelism and Dual-scope attention to significantly accelerate long-form video generation across multiple GPUs.
Alexander Gambashidze, Anton Kulikov, Yuriy Sosnin +1 authors
Using perceptual objectives in U-Net embedding space for text-to-image diffusion models enhances human preference alignment, visual appeal, and efficiency compared to traditional methods.
Zuxin Liu, Thai Hoang, Jianguo Zhang +14 authors
APIGen is an automated pipeline for generating high-quality, verified datasets for function-calling applications, demonstrating superior performance in the Berkeley Function-Calling Benchmark.
Yuxuan Wang, Yueqian Wang, Dongyan Zhao +2 authors
VideoHallucer, a comprehensive benchmark for detecting hallucinations in video-language models, reveals significant issues with current models and instructs the development of a self-PEP framework for improved hallucination resistance.
Yantao Liu, Zhao Zhang, Zijun Yao +3 authors
ARTE, a framework that aligns teacher models with student preferences, enhances knowledge distillation for edge deployment of LLMs, demonstrating superior performance and generalization compared to existing methods.
Jiahe Huang, Guandao Yang, Zichen Wang +1 authors
DiffusionPDE, a generative diffusion model framework, solves PDEs effectively under partial observations by modeling the joint distribution of solutions and coefficients.
Shawn Gavin, Tuney Zheng, Jiaheng Liu +7 authors
The LongIns benchmark evaluates large language models' long-context reasoning abilities, revealing performance issues even in models with large context windows.
Alexandre Ramé, Johan Ferret, Nino Vieillard +7 authors
A new alignment strategy named Weight Averaged Rewarded Policies (WARP) enhances reinforcement learning from human feedback by merging policies at multiple stages to balance KL regularization and reward optimization, improving large language model quality and alignment.
Xuyang Shen, Dong Li, Ruitao Leng +3 authors
Linear complexity language models show similar scalability and superior linguistic performance compared to traditional transformer-based models.
Colin White, Samuel Dooley, Manley Roberts +12 authors
LiveBench is a benchmark designed to be contamination-free and bias-resilient, featuring automatically scored questions from various sources and encompassing diverse tasks including math, coding, reasoning, and data analysis.
Yannis Tevissen, Khalil Guetari, Frédéric Petitpont
The proposed system uses Retrieval Augmented Generation (RAG) and large language models to efficiently answer questions about video libraries by generating relevant search queries, retrieving video moments, and producing timestamped responses.
Xuefei Ning, Zifu Wang, Shiyao Li +7 authors
LLMs can improve through a learning-by-teaching approach, potentially leading to self-improvement without relying solely on human data or more powerful models.
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