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。
An Yang, Baosong Yang, Binyuan Hui +55 authors
The Qwen2 series, comprising 0.5 to 72 billion parameter models, surpasses prior open models across language understanding, generation, multilingualism, coding, math, and reasoning, with exceptional performance in benchmarks like MMLU, GPQA, HumanEval, GSM8K, BBH, MT-Bench, Arena-Hard, and LiveCodeBench.
50 篇论文 · 按点赞排序
Yuzhang Tian, Jianbo Zhao, Haoyu Dong +8 authors
SpreadsheetLLM introduces SheetCompressor and Chain of Spreadsheet to enhance LLMs' performance on spreadsheet tasks through efficient encoding and understanding.
Ayush Kaushal, Tejas Pandey, Tejas Vaidhya +2 authors
The Spectra LLM suite provides insights into the performance and characteristics of low-bitwidth models through comprehensive training and evaluation of FloatLMs, QuantLMs, and TriLMs across various benchmarks and datasets.
Yunfei Chu, Jin Xu, Qian Yang +9 authors
Qwen2-Audio, a large-scale audio-language model, enhances instruction-following and audio analysis through natural language prompts and DPO optimization.
Zafeirios Fountas, Martin A Benfeghoul, Adnan Oomerjee +4 authors
EM-LLM enhances LLMs by integrating human episodic memory principles, enabling efficient handling of extensive contexts through coherent event segmentation and retrieval.
Daniel Goldstein, Fares Obeid, Eric Alcaide +2 authors
A hybrid Linear Attention/Transformer model, GoldFinch, efficiently generates a compressed KV-Cache for large context lengths with linear time and space complexity, outperforming previous models.
Chaofan Tao, Qian Liu, Longxu Dou +5 authors
Investigating the impact of vocabulary size on the scaling of large language models reveals that larger vocabularies improve performance when considering compute budgets, and current models often use suboptimal vocabulary sizes.
Zhaorun Chen, Zhen Xiang, Chaowei Xiao +2 authors
AgentPoison, a novel red teaming approach, introduces a backdoor attack targeting LLM agents by poisoning their memory or knowledge base, ensuring high attack success rates with minimal impact on normal performance.
Ting Jiang, Minghui Song, Zihan Zhang +6 authors
E5-V, a novel framework, utilizes multimodal large language models (MLLMs) to achieve universal multimodal embeddings with strong performance and reduced costs through single modality training.
Mo Li, Songyang Zhang, Yunxin Liu +1 authors
NeedleBench and the Ancestral Trace Challenge evaluate the long-context retrieval and reasoning capabilities of large language models, highlighting their limitations in practical applications.
Kaichen Zhang, Bo Li, Peiyuan Zhang +8 authors
LMMS-EVAL and LMMS-EVAL LITE provide frameworks for evaluating large multi-modal models with comprehensive coverage, while Multimodal LIVEBENCH assesses models' generalization using real-world data.
Ben Cohen, Emaad Khwaja, Kan Wang +4 authors
Toto, a Time Series Optimized Transformer for Observability, achieves state-of-the-art performance in observability and general-purpose forecasting using a vast dataset of time series data.
Rulin Shao, Jacqueline He, Akari Asai +5 authors
Increasing the size of the datastore for retrieval-based language models improves performance without obvious saturation, suggesting datastore size is a key factor in model efficiency and performance.
Zhenhua Liu, Tong Zhu, Chuanyuan Tan +1 authors
The Name-Aware Unlearning Framework for Privacy Protection demonstrates superior capabilities in protecting individual private data in large language models without needing complete retraining.
Zhengcong Fei, Mingyuan Fan, Changqian Yu +2 authors
DiT-MoE, a sparse diffusion Transformer with shared expert routing and expert-level balance loss, achieves competitive performance with dense networks in image generation while reducing computational load during inference.
Hanqing Wang, Jiahe Chen, Wensi Huang +19 authors
Project GRUtopia introduces GRScenes, GRResidents, and GRBench to simulate a comprehensive 3D interactive society for embodied AI, focusing on diverse environments and advanced robotic tasks.
Yaoting Wang, Peiwen Sun, Dongzhan Zhou +3 authors
A new task, Reference Audio-Visual Segmentation (Ref-AVS), is introduced to segment objects using multimodal cues, and a method leveraging these cues outperforms existing approaches in experiments.
Hongyu Wang, Shuming Ma, Ruiping Wang +1 authors
Q-Sparse enhances sparsely-activated large language models for efficient inference across various training and deployment scenarios, using top-K sparsification and straight-through-estimators.
Yifan Song, Guoyin Wang, Sujian Li +1 authors
Research explores non-determinism in LLM evaluations, comparing greedy decoding and sampling, and finds that smaller models can match larger ones with best-of-N sampling.
Jessica Echterhoff, Fartash Faghri, Raviteja Vemulapalli +4 authors
The work provides evaluation metrics and a training strategy to minimize inconsistencies and negative flips in Large Language Model updates, ensuring better compatibility with prior model versions for downstream tasks.
Matt Stallone, Vaibhav Saxena, Leonid Karlinsky +19 authors
Granite code models with extended context windows up to 128K tokens are achieved through lightweight continual pretraining and finetuning with increased RoPE frequencies and length-upsampled data, showing improvements in long-context tasks without degrading performance on standard benchmarks.
Qianqian Wang, Vickie Ye, Hang Gao +3 authors
A monocular dynamic reconstruction method using SE3 motion bases and data-driven priors achieves state-of-the-art performance in 3D/2D motion estimation and novel view synthesis.
Huanqian Wang, Yang Yue, Rui Lu +5 authors
Directly editing a subset of LLM parameters can effectively modulate specific behaviors like detoxification and jailbreak resistance with minimal computational cost while maintaining general LLM capabilities.
Pascal Pfeiffer, Philipp Singer, Yauhen Babakhin +3 authors
H2O-Danube3, a series of small language models, achieves high performance across various benchmarks and is efficient for local inference on smartphones.
Boyang Deng, Richard Tucker, Zhengqi Li +3 authors
A method synthesizes long-range Streetscapes video sequences from city layouts and language inputs using conditioned video diffusion within an autoregressive framework.
Yichi Zhang, Yao Huang, Yitong Sun +10 authors
MultiTrust is a comprehensive benchmark assessing the trustworthiness of MLLMs across truthfulness, safety, robustness, fairness, and privacy, revealing multimodal challenges and the need for advanced methodologies.
Yixin Liu, Pengfei Liu, Arman Cohan
Direct Preference Optimization (DPO) is sensitive to KL-divergence constraints and benefits from reference policies similar to the model being fine-tuned, offering insights into best practices and future research.
Chenze Shao, Fandong Meng, Jie Zhou
Patch-level training for LLMs reduces computational costs by compressing multiple tokens into patches, decreasing sequence length while maintaining model performance.
Haodong Duan, Junming Yang, Yuxuan Qiao +9 authors
VLMEvalKit is a PyTorch-based toolkit for evaluating large multi-modality models with a focus on vision-language models, offering a comprehensive framework for reproducible results and a leaderboard for research tracking.
Yulong Wang, Tianhao Shen, Lifeng Liu +1 authors
Sibyl enhances complex reasoning by integrating a global workspace and multi-agent debate, improving LLM-based agents' performance on advanced 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