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。
Andrew Zhao, Yiran Wu, Yang Yue +8 authors
Absolute Zero Reasoner (AZR) achieves state-of-the-art performance on coding and mathematical reasoning tasks through self-generated, verified learning without external data.
50 篇论文 · 按点赞排序
Yunxin Li, Zhenyu Liu, Zitao Li +19 authors
This survey outlines the development of multimodal reasoning models, progressing from task-specific modules to unified frameworks, and discusses emerging capabilities and challenges in integrating reasoning across diverse data types.
Roman Abramov, Felix Steinbauer, Gjergji Kasneci
Grokking with synthetic data augmentation in knowledge graphs enhances Transformer models' multi-step factual reasoning on real-world tasks, achieving high accuracy on benchmarks.
Yibin Wang, Zhimin Li, Yuhang Zang +4 authors
A unified multimodal reward model that incorporates long chains of thought reasoning improves reliability and accuracy in vision reward tasks through exploration-driven reinforcement learning.
Yemin Shi, Yu Shu, Siwei Dong +4 authors
Voila is an end-to-end voice-language model that enables low-latency, emotionally expressive voice interactions and supports multiple applications including ASR, TTS, and multilingual speech translation.
Hao Fei, Yuan Zhou, Juncheng Li +29 authors
A General-Level evaluation framework is introduced to assess multimodal generalist performance in LLMs, using Synergy to measure consistent capabilities across comprehension and generation.
Xinjie Zhang, Jintao Guo, Shanshan Zhao +7 authors
A survey examines the integration of multimodal understanding and image generation models, categorizing them into diffusion-based, autoregressive-based, and hybrid approaches, and discussing challenges and future research directions.
Xiusi Chen, Gaotang Li, Ziqi Wang +9 authors
Reasoning Reward Models (ReasRMs) enhance reward modeling for large language models by integrating reasoning tasks, improving interpretability and performance.
Jie Liu, Gongye Liu, Jiajun Liang +6 authors
Flow-GRPO combines online reinforcement learning with flow matching models through an ODE-to-SDE conversion and denoising reduction, improving sampling efficiency and performance across text-to-image tasks.
Hao Sun, Zile Qiao, Jiayan Guo +6 authors
ZeroSearch is a reinforcement learning framework that improves LLM search capabilities without using real search engines, using supervised fine-tuning and curriculum-based rollouts to handle document quality and API cost challenges.
Ziyang Xu, Kangsheng Duan, Xiaolei Shen +5 authors
PixelHacker, a diffusion-based model using latent categories guidance and linear attention, enhances image inpainting by improving structural and semantic consistency across various datasets.
Akhiad Bercovich, Itay Levy, Izik Golan +129 authors
Llama-Nemotron models offer exceptional reasoning capabilities, inference efficiency, and open licensing through neural architecture search, knowledge distillation, and reasoning-focused post-training, including supervised fine-tuning and reinforcement learning.
Sihyeong Park, Sungryeol Jeon, Chaelyn Lee +3 authors
A comprehensive evaluation of 25 open-source and commercial LLM inference engines across various criteria and optimization techniques is provided, with an outline of future research directions.
Joykirat Singh, Raghav Magazine, Yash Pandya +1 authors
ARTIST integrates agentic reasoning and reinforcement learning to enhance LLMs' ability to dynamically use tools and interact with environments, significantly improving performance on complex reasoning tasks.
Teng Hu, Zhentao Yu, Zhengguang Zhou +4 authors
HunyuanCustom is a multi-modal video generation framework that enhances identity consistency and handles various types of input conditions through specific fusion and injection mechanisms, outperforming existing methods.
Daniel Goldstein, Eric Alcaide, Janna Lu +1 authors
A protocol converts softmax attention transformers into linear attention decoders using minimal tokens while maintaining quality and performance.
Zhouliang Yu, Ruotian Peng, Keyi Ding +10 authors
FormalMATH, a large-scale Lean4 benchmark, introduces an autoformalization pipeline to reduce manual annotation costs and identifies limitations of existing LLM-based theorem provers in formal mathematical reasoning.
Essential AI, Ishaan Shah, Anthony M. Polloreno +21 authors
Muon, a second-order optimizer, improves data efficiency and computational savings over AdamW, especially at large batch sizes, and combined with muP, it provides efficient hyperparameter transfer and minimal resource overhead.
Ava Pun, Kangle Deng, Ruixuan Liu +3 authors
LegoGPT generates physically stable LEGO models from text prompts using a large language model with physics-aware constraints and supports manual and automated assembly.
Bang Zhang, Ruotian Ma, Qingxuan Jiang +10 authors
SAGE, an automated evaluation framework, assesses the higher-order social cognition and empathy of large language models through simulated human-like emotions and inner thoughts in multi-turn conversations.
Xianhang Li, Yanqing Liu, Haoqin Tu +2 authors
OpenVision provides a fully-open, cost-effective family of vision encoders that matches or surpasses OpenAI's CLIP in multimodal frameworks, offering flexible trade-offs between model size and performance.
Yaoqi Chen, Jinkai Zhang, Baotong Lu +15 authors
RetroInfer, a system that uses a vector storage approach based on attention sparsity, significantly accelerates inference for large language models with long contexts without reducing accuracy.
Daneul Kim, Jaeah Lee, Jaesik Park
A framework for iterative image editing uses layer-wise memory and background consistency guidance to maintain scene coherence and adapt new elements naturally across sequential edits.
Shiyi Zhang, Junhao Zhuang, Zhaoyang Zhang +2 authors
FlexiAct transfers actions from reference videos to arbitrary target images, adapting to diverse layouts, viewpoints, and skeletal structures while maintaining identity consistency using RefAdapter and FAE.
Yi-Fan Zhang, Xingyu Lu, Xiao Hu +13 authors
Using Reinforcement Learning to improve reward modeling in Multimodal Reward Models leads to more stable training and superior performance.
Jingwen Ye, Yuze He, Yanning Zhou +5 authors
PrimitiveAnything is a novel framework for shape primitive abstraction that reformulates the task as primitive assembly generation using a shape-conditioned primitive transformer and achieves high-quality results across diverse shape categories.
Yuhui Xu, Hanze Dong, Lei Wang +3 authors
Elastic Reasoning is a framework that divides reasoning into thinking and solution phases with separate budgets, improving model reliability and efficiency under resource constraints.
Albert Ge, Tzu-Heng Huang, John Cooper +7 authors
R&B, a framework that repartitions and balances training data based on semantic similarity and domain gradients, enhances language model performance with minimal additional computational cost.
Hanhua Hong, Chenghao Xiao, Yang Wang +3 authors
Inversion learning automates the generation of effective evaluation prompts for language models, improving robustness and efficiency over manual processes.
Gracjan Góral, Alicja Ziarko, Piotr Miłoś +3 authors
State-of-the-art Vision Language Models excel in scene understanding but struggle with spatial reasoning and visual perspective taking in controlled visual 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