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。
50 篇论文 · 按点赞排序
Wei He, Zhiheng Xi, Wanxu Zhao +6 authors
A method called Code-as-Intermediary Translation (CIT) uses code to translate visual charts into text, enhancing MLLMs with visual reasoning abilities through cost-effective data synthesis.
Yu Zhao, Alessio Devoto, Giwon Hong +5 authors
A training-free method called SpARE uses pre-trained sparse auto-encoders to control knowledge selection in LLMs, effectively resolving context-memory knowledge conflicts in open-domain QA tasks.
Yiran Qin, Zhelun Shi, Jiwen Yu +10 authors
A dual evaluation framework, WorldSimBench, is proposed to assess World Simulators from both visual and action perspectives using explicit perceptual evaluation and implicit manipulative evaluation in embodied scenarios.
Hao Peng, Xin Lv, Yushi Bai +4 authors
The study extends knowledge distillation to the pre-training phase of large language models, demonstrating its effectiveness and exploring key design factors like logits processing, loss selection, scaling laws, and offline/online logits.
Zhifei Xie, Changqiao Wu
Mini-Omni2 is a real-time, end-to-end voice response model that integrates visual and auditory encoders and employs a three-stage training process to handle multi-modal data.
Chien Van Nguyen, Huy Huu Nguyen, Thang M. Pham +8 authors
Taipan combines Mamba-2 with Selective Attention Layers to achieve efficient long-context language modeling with high performance and constant memory usage.
Michael S. Ryoo, Honglu Zhou, Shrikant Kendre +7 authors
BLIP-3-Video is a video-specific multimodal language model using fewer visual tokens and a temporal encoder for efficient temporal information capture and accurate video question-answering.
Yun Xing, Yiheng Li, Ivan Laptev +1 authors
Concentric Causal Attention (CCA) mitigates object hallucination in Large Vision Language Models by addressing long-term decay in Rotary Position Encoding (RoPE) during multimodal interactions.
Chenxin An, Jun Zhang, Ming Zhong +5 authors
STRING, a position embedding technique that shifts well-trained positions during inference, enhances the performance of large language models without additional training, surpassing both open-source and commercial models on long-context benchmarks.
Shansan Gong, Shivam Agarwal, Yizhe Zhang +9 authors
Converted autoregressive language models into diffusion language models, demonstrating competitive performance on language modeling and reasoning tasks using fewer training tokens.
Matthieu Lin, Jenny Sheng, Andrew Zhao +7 authors
A survey of principles and trends in using large language models (LLMs) for end-to-end optimization of compound AI systems, focusing on LLM-driven approaches and program analysis concepts.
Yuxian Gu, Hao Zhou, Fandong Meng +2 authors
MiniPLM is a knowledge distillation framework that enhances pre-training efficiency, flexibility, and effectiveness for student language models by refining the training data distribution using offline teacher inference.
Haotian Tang, Yecheng Wu, Shang Yang +7 authors
HART, a hybrid autoregressive transformers model, generates 1024x1024 images with high quality and efficiency by combining discrete and continuous tokenization.
Yijun Liang, Shweta Bhardwaj, Tianyi Zhou
A novel diffusion curriculum dynamically adjusts image guidance to improve learning from low-quality or scarce data, enhancing model accuracy on both in-distribution and out-of-distribution tasks.
Ling-Hao Chen, Wenxun Dai, Xuan Ju +2 authors
MotionCLR, an attention-based motion diffusion model, enhances interactive motion editing by explicitly modeling word-level text-motion correspondence and improving explainability through self-attention and cross-attention mechanisms.
Shyamgopal Karthik, Huseyin Coskun, Zeynep Akata +3 authors
A scalable approach using a pre-trained reward function generates synthetic preference datasets for Direct Preference Optimization (DPO) in text-to-image models, improving prompt-following and visual quality.
Shota Onohara, Atsuyuki Miyai, Yuki Imajuku +5 authors
JMMMU, a Japanese benchmark for Large Multimodal Models, evaluates their performance on expert-level tasks and cultural understanding through culture-agnostic and culture-specific subsets.
Amirhossein Alimohammadi, Sauradip Nag, Saeid Asgari Taghanaki +3 authors
A pre-trained text-to-image diffusion model with tracking mechanism outperforms existing methods in object segmentation across video frames with arbitrary granularity.
Hengwei Bian, Lingdong Kong, Haozhe Xie +3 authors
DynamicCity generates large-scale, high-quality 4D LiDAR scenes by integrating a VAE model with a DiT-based diffusion model, effectively capturing dynamic environments.
Shaonan Wu, Shuai Lu, Yeyun Gong +2 authors
Alchemy synthesizes formal theorems through symbolic mutation to augment Mathlib, enhancing performance on theorem proving benchmarks.
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