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 Gorbatovski, Boris Shaposhnikov, Alexey Malakhov +5 authors
A new method, Trust Region DPO (TR-DPO), is proposed to improve policy alignment in reinforcement learning, outperforming Direct Preference Optimization (DPO) by up to 19% on key datasets by updating the reference policy during training.
33 篇论文 · 按点赞排序
Xuezhe Ma, Xiaomeng Yang, Wenhan Xiong +7 authors
Megaldon, an enhanced neural architecture, provides efficient long-sequence modeling by combining and improving upon components from mega and transformer models, achieving better pretraining efficiency and performance.
Ye Tian, Baolin Peng, Linfeng Song +4 authors
AlphaLLM enhances Large Language Models through integration with Monte Carlo Tree Search and critic models, improving performance in complex reasoning tasks without additional annotations.
Dongseong Hwang, Weiran Wang, Zhuoyuan Huo +2 authors
Feedback Attention Memory (FAM) enhances Transformer architecture by enabling long-context processing without additional weights, significantly improving performance on large sequences across various model sizes.
Aitor Ormazabal, Che Zheng, Cyprien de Masson d'Autume +22 authors
Reka Core, Flash, and Edge are multimodal language models that achieve state-of-the-art performance across text, image, video, and audio tasks, offering high efficiency and competitive results relative to their compute class.
Zichen Liu, Yihao Meng, Hao Ouyang +4 authors
Dynamic Typography generates coherent and semantically meaningful text animations using neural displacement fields, end-to-end optimization, and perceptual loss regularization, outperforming baseline methods.
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.
Hongchi Xia, Zhi-Hao Lin, Wei-Chiu Ma +1 authors
Video2Game converts real-world scenes into realistic interactive environments by combining neural radiance fields, mesh distillation, and physics modeling.
Zichao Li, Cihang Xie, Ekin Dogus Cubuk
Research explores the performance of scaled-down CLIP models, examining data quality, model architecture, and training strategies to optimize resource utilization and performance.
Xueqing Deng, Qihang Yu, Peng Wang +2 authors
COCONut, a reevaluated and expanded version of the COCO dataset, provides high-quality annotations for semantic, instance, and panoptic segmentation tasks, serving as the first large-scale universal segmentation benchmark.
SIMA Team, Maria Abi Raad, Arun Ahuja +88 authors
Developing a scalable and adaptable embodied AI agent capable of following complex language instructions across various 3D environments using minimal assumptions.
Yuzhen Huang, Jinghan Zhang, Zifei Shan +1 authors
Research shows that the ability of large language models to compress text is linearly associated with their performance on various benchmarks related to intelligence, suggesting compression efficiency as a reliable evaluation metric.
Xinyue Wei, Kai Zhang, Sai Bi +6 authors
MeshLRM, an LRM-based approach, reconstructs high-quality meshes efficiently from four images using differentiable mesh extraction and rendering, enhancing NeRF-based methods.
Xingyu Fu, Yushi Hu, Bangzheng Li +7 authors
Blink is a benchmark for evaluating the visual perception abilities of multimodal language models, demonstrating current limitations compared to human performance and specialist computer vision models.
Zach Evans, Julian D. Parker, CJ Carr +3 authors
A diffusion-transformer trained on long temporal contexts can generate full-length music tracks with coherent structure at a latent rate of 21.5Hz.
Thibault Castells, Hyoung-Kyu Song, Tairen Piao +6 authors
Researchers achieve efficient text-to-image generation with reduced sampling steps and architectural optimizations, specifically using a compact Stable Diffusion variant, high-quality datasets, and an advanced distillation process for rapid and accurate image generation on edge devices.
Han Lin, Jaemin Cho, Abhay Zala +1 authors
Ctrl-Adapter efficiently integrates pretrained ControlNets into various diffusion models to enable versatile image and video control with superior performance and reduced computational costs.
Hanshi Sun, Zhuoming Chen, Xinyu Yang +2 authors
TriForce, a hierarchical speculative decoding system, accelerates long-sequence inference for large language models by using a dynamic sparse KV cache and layered model speculation, achieving significant speedups compared to baseline methods.
Zhaofeng Wu, Ananth Balashankar, Yoon Kim +2 authors
A zero-shot cross-lingual alignment method for language models using a reward model trained in one language outperforms unaligned models across multiple languages and tasks.
Kuan-Chieh, Wang, Daniil Ostashev +3 authors
Mixture-of-Attention (MoA) architecture personalizes text-to-image diffusion models by blending a fixed attention prior branch with a learnable personalized branch to enhance subject-context control.
Mude Hui, Siwei Yang, Bingchen Zhao +5 authors
HQ-Edit, a high-quality instruction-based image editing dataset using GPT-4V and DALL-E 3, enhances image editing models with precise text prompts and evaluation metrics.
Mohamed El Banani, Amit Raj, Kevis-Kokitsi Maninis +7 authors
Visual foundation models are analyzed for their 3D awareness through task-specific probes and zero-shot inference on frozen features, revealing limitations.
Bertie Vidgen, Adarsh Agrawal, Ahmed M. Ahmed +94 authors
The AI Safety Benchmark v0.5, developed by MLCommons, assesses the safety risks of chat-tuned language models using a taxonomy of 13 hazard categories and includes 43,090 test items.
Ronghuan Wu, Wanchao Su, Kede Ma +1 authors
AniClipart system transforms static clipart images into high-quality motion sequences using Bézier curves, text-to-video priors, and Video Score Distillation Sampling loss for coherent and identity-preserved animations.
Mukul Gagrani, Raghavv Goel, Wonseok Jeon +3 authors
Speculative decoding enhances the inference efficiency of Multimodal Large Language Models (MLLMs), specifically LLaVA 7B, by using a language-only draft model and a compact LLaVA model with an image adapter.
Chandeepa Dissanayake, Lahiru Lowe, Sachith Gunasekara +1 authors
A fine-tuning recipe using OpenLLaMA 3Bv2, synthetic data generation, QLoRA, HH-RLHF, and DPO yields superior performance for the OpenBezoar family of models.
Agneet Chatterjee, Tejas Gokhale, Chitta Baral +1 authors
Recent monocular depth estimation techniques using language guidance show optimal performance with scene-level descriptions but degrade with low-level details and adversarial attacks.
Ya-Qi Yu, Minghui Liao, Jihao Wu +3 authors
TextHawk, a multimodal large language model designed for document-oriented tasks, outperforms existing models by using specialized components for efficient text and image perception.
Navonil Majumder, Chia-Yu Hung, Deepanway Ghosal +3 authors
Improving text-to-audio generation by fine-tuning the Tango model using diffusion-DPO on a preference dataset enhances audio quality in terms of automatic and manual evaluations.
Jonathan D. Chang, Wenhao Shan, Owen Oertell +4 authors
The DR-PO algorithm leverages reset techniques in RLHF to improve policy optimization by incorporating an offline preference dataset, resulting in better generative performance compared to PPO and DPO.
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