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
Amirhosein Ghasemabadi, Keith G. Mills, Baochun Li +1 authors
The Guided by Gut (GG) framework enhances LLM reasoning efficiently using intrinsic signals and token-level confidence, outperforming PRM-based methods with faster inference and lower memory usage.
Yi Wang, Junxiao Liu, Shimao Zhang +2 authors
PATS enhances LLM efficiency by dynamically adjusting reasoning strategies based on task difficulty, leveraging PRMs and Beam Search.
Jiangjie Chen, Qianyu He, Siyu Yuan +9 authors
Enigmata is a comprehensive suite for improving LLMs in puzzle reasoning through scalable multi-task RL training, leading to better performance on benchmarks and advanced math tasks.
Chenyu Yang, Shiqian Su, Shi Liu +11 authors
ZeroGUI is an online learning framework that uses Vision-Language Models for task generation and reward estimation, enhancing GUI Agents' performance with minimal human intervention.
Jintao Zhang, Xiaoming Xu, Jia Wei +5 authors
SageAttention2++ improves attention efficiency by using FP8 Matmul in FP16, achieving a 3.9x speedup over FlashAttention without losing accuracy.
Siye Wu, Jian Xie, Yikai Zhang +4 authors
Adaptive Reasoning Model (ARM) uses Ada-GRPO to reduce token usage and improve efficiency across different reasoning modes.
Guiyao Tie, Xueyang Zhou, Tianhe Gu +7 authors
The MMMR benchmark evaluates multi-modal reasoning in MLLMs by assessing thinking quality through diverse reasoning types and a modular evaluation pipeline.
Ang Lv, Ruobing Xie, Xingwu Sun +2 authors
LLMs exhibit robustness to reward noise during post-training and achieve high performance using reasoning pattern rewards (RPR) in conjunction with noisy reward models.
Yi Ding, Ruqi Zhang
Sherlock, a self-correction and self-improvement framework for reasoning vision-language models, enhances accuracy across benchmarks using limited annotated data.
Weizhou Shen, Chenliang Li, Fanqi Wan +12 authors
QwenLong-CPRS enhances large language models with multi-granularity context compression, dynamic optimization guided by natural language, and efficient bidirectional reasoning and parallel inference, achieving superior performance and context management.
Yunxin Li, Xinyu Chen, Zitao Li +5 authors
A Verifier-guided Iterative Policy Optimization method enhances Video-LLMs' reasoning capabilities by integrating a Rollout-Aware Verifier between GRPO and DPO phases, leading to faster and more effective optimization.
Shuo Yang, Haocheng Xi, Yilong Zhao +11 authors
SVG2 is a training-free framework that enhances video generation efficiency and quality by accurately identifying and processing critical tokens using semantic-aware permutation and dynamic budget control.
Haoran He, Jiajun Liang, Xintao Wang +4 authors
EvoSearch, an evolutionary search method, enhances test-time scaling for diffusion and flow-based generative models, improving image and video generation quality, diversity, and generalizability.
Maksim Kolodiazhnyi, Denis Tarasov, Dmitrii Zhemchuzhnikov +6 authors
A multi-modal CAD reconstruction model leveraging vision-language models and reinforcement learning achieves state-of-the-art performance across various datasets, including real-world ones.
Qiushi Sun, Zhoumianze Liu, Chang Ma +18 authors
ScienceBoard provides a realistic scientific workflow environment and benchmark to evaluate the performance of LLM-based agents, demonstrating their current limitations in complex scientific tasks.
Yuanxin Liu, Kun Ouyang, Haoning Wu +7 authors
A new benchmark, VideoReasonBench, evaluates complex vision-centric video reasoning, finding that extended thinking budgets are crucial for improved performance compared to existing benchmarks.
Han Xiao, Guozhi Wang, Yuxiang Chai +12 authors
UI-Genie framework addresses GUI agent challenges through a reward model with image-text architecture and a self-improvement pipeline, achieving state-of-the-art performance on multiple benchmarks.
Yang Shi, Huanqian Wang, Wulin Xie +15 authors
MLLMs achieve modest accuracy in video OCR due to motion blur, temporal variations, and visual effects; MME-VideoOCR benchmark reveals limitations in spatio-temporal reasoning and language bias.
Chong Zeng, Yue Dong, Pieter Peers +2 authors
RenderFormer is a transformer-based neural rendering pipeline that renders images from triangle representations without per-scene training and with full global illumination effects.
Yeonjoon Jung, Daehyun Ahn, Hyungjun Kim +2 authors
Granular Low-Rank Adaptation (GraLoRA) improves upon Low-Rank Adaptation (LoRA) by partitioning weight matrices to mitigate overfitting and enhance performance in parameter-efficient fine-tuning.
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