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。
Alexey Dontsov, Dmitrii Korzh, Alexey Zhavoronkin +6 authors
CLEAR benchmark evaluates multimodal unlearning methods across textual and visual data, highlighting challenges and demonstrating the effectiveness of $\ell_1$ regularization on LoRA weights in mitigating catastrophic forgetting.
50 篇论文 · 按点赞排序
OpenAI, Aaron Hurst, Adam Lerer +416 authors
GPT-4o is an omnimodal autoregressive model trained to handle text, audio, image, and video inputs, offering high-performance outputs across these modalities, with particular strengths in vision and audio.
Viacheslav Surkov, Chris Wendler, Mikhail Terekhov +3 authors
Sparse autoencoders are applied to learn interpretable features from intermediate updates in text-to-image diffusion models, such as SDXL Turbo, enhancing understanding of these models' internal mechanisms.
Ming Li, Yanhong Li, Tianyi Zhou
Investigating gradient patterns in LLMs during fast vs. slow training reveals that slow thinking (chain-of-thoughts) enhances learning stability and distinguishes correct reasoning paths.
Shaofei Cai, Zihao Wang, Kewei Lian +4 authors
Visual-temporal context prompting integrates object segmentation and visual observations to enhance VLMs in embodied decision-making, enabling complex spatial tasks.
Yiruo Cheng, Kelong Mao, Ziliang Zhao +6 authors
CORAL is a large-scale benchmark for evaluating Retrieval-Augmented Generation (RAG) systems in multi-turn conversational settings, addressing challenges like open-domain coverage, knowledge intensity, free-form responses, and topic shifts.
Chien Van Nguyen, Xuan Shen, Ryan Aponte +25 authors
A survey of small language models covers their architectures, training methods, and model compression techniques, with a taxonomy for optimization, benchmark datasets, evaluation metrics, and open challenges.
Ziming Li, Qianbo Zang, David Ma +11 authors
AutoKaggle, a collaborative multi-agent system, enhances data science productivity through an iterative development process that combines code execution, debugging, and testing with human expertise and automated tools.
Chengyou Jia, Minnan Luo, Zhuohang Dang +5 authors
AgentStore integrates heterogeneous agents using MetaAgent and AgentToken strategies, enhancing generalization and specialization in computer tasks, especially outperforming previous systems on the OSWorld benchmark.
Qintong Zhang, Victor Shea-Jay Huang, Bin Wang +7 authors
The survey reviews document parsing methodologies, including modular pipeline systems and end-to-end models, focusing on layout detection, content extraction, and multi-modal data integration, while highlighting challenges and future research directions.
Arnon Turetzky, Nimrod Shabtay, Slava Shechtman +4 authors
SALAD, a per-token latent diffusion model using continuous representations, achieves superior intelligibility in zero-shot text-to-speech without compromising speech quality and speaker similarity.
Ruoqi Liu, Yuelin Bai, Xiang Yue +1 authors
ECGInstruct and ECGBench improve ECG image interpretation using a specialized multimodal large language model, PULSE, which outperforms general models.
Ankan Mullick, Sombit Bose, Abhilash Nandy +2 authors
A pointer network-based architecture improves multi-intent and multi-span intent detection in task-oriented dialogue systems using a new multi-lingual dataset.
Krzysztof Ociepa, Łukasz Flis, Krzysztof Wróbel +2 authors
Bielik 7B v0.1, a 7-billion-parameter generative text model for Polish, uses Weighted Instruction Cross-Entropy Loss and Adaptive Learning Rate to enhance performance in various NLP tasks, surpassing Mistral-7B-v0.1 and setting new benchmarks.
Haiyang Wang, Yue Fan, Muhammad Ferjad Naeem +5 authors
TokenFormer, a scalable architecture for Transformers, enhances flexibility by treating model parameters as tokens, allowing for efficient scaling without retraining.
Zhengyao Lv, Chenyang Si, Junhao Song +4 authors
FasterCache, a training-free strategy, accelerates video diffusion model inference by dynamically reusing features while maintaining video quality.
S Sakshi, Utkarsh Tyagi, Sonal Kumar +6 authors
MMAU, a new benchmark for evaluating advanced audio understanding models, comprises audio clips with complex reasoning tasks and requires models to demonstrate specialized skills, highlighting significant areas for improvement in current models.
Yuxiang Wei, Federico Cassano, Jiawei Liu +7 authors
SelfCodeAlign is a transparent and self-aligning pipeline for enhancing large language models' ability to follow human instructions without human annotations or distillation, significantly improving coding performance across various benchmarks.
Haozhe Liu, Shikun Liu, Zijian Zhou +12 authors
MarDini combines masked auto-regression and diffusion modeling for efficient video generation, accommodating various scenarios like interpolation and expansion with high quality.
Thomas Schmied, Thomas Adler, Vihang Patil +6 authors
A Large Recurrent Action Model (LRAM) using xLSTM architecture achieves comparable performance and faster inference speed than Transformer-based models for large action models.
Xinghao Wang, Pengyu Wang, Bo Wang +3 authors
BitStack, a training-free weight compression approach, dynamically adjusts model size and memory usage with fine-grained control, outperforming traditional quantization methods.
Haocheng Xi, Han Cai, Ligeng Zhu +4 authors
COAT reduces memory usage and accelerates training of large models through dynamic range expansion and mixed-granularity activation quantization in FP8 training.
Yuang Ai, Xiaoqiang Zhou, Huaibo Huang +4 authors
A dual strategy involving a data curation pipeline (GenIR) and a DiT-based model (DreamClear) enhances image restoration by creating a large dataset and utilizing generative priors and adaptive modulators.
Jiajie Zhang, Zhongni Hou, Xin Lv +7 authors
A novel reward-based method called LongReward enhances long-context performance of supervised fine-tuned models by leveraging human-valued criteria and an off-the-shelf LLM, and it can be combined with offline RL algorithms like DPO effectively.
Yunjia Qi, Hao Peng, Xiaozhi Wang +3 authors
Constraint back-translation enhances large language models' ability to follow complex instructions by leveraging implicit constraints in existing datasets, improving their performance on instruction-following benchmarks.
Shanchao Liang, Yiran Hu, Nan Jiang +1 authors
REPOCOD, a new code generation benchmark using real-world software development problems, shows that existing LLMs do not perform well and highlights the need for more robust models to assist developers.
Hongliang He, Wenlin Yao, Kaixin Ma +5 authors
An open-source framework enables the development of a multimodal web agent that autonomously explores and improves through an imitation learning and exploration-feedback-optimization cycle.
Wanhua Li, Zibin Meng, Jiawei Zhou +3 authors
A modular framework that leverages Vision Foundation Models and Large Language Models for social relation reasoning achieves competitive zero-shot accuracy and interpretability through text-based explanations, with automatic prompt optimization to enhance performance.
Yihe Deng, Paul Mineiro
A novel approach using online learning Flows with Direct Preference Optimization enhances the quality of reasoning traces for LLMs in mathematical reasoning tasks.
Shuhao Gu, Jialing Zhang, Siyuan Zhou +16 authors
Infinity-MM, a large-scale multimodal instruction dataset, enhances open-source VLMs through rigorous filtering and synthetic data generation, achieving state-of-the-art performance in models of similar size.
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