arXivDaily arXiv每日学术速递 周一至周五更新

高校专区

University of Southern California(南加州大学)

2026-08-05 至 2026-08-05 共收录 2
2607.27670 2026-08-05 cs.CV cs.AI 版本更新

JigShape: Evaluating Visual-Geometric Reasoning in VLMs through Jigsaw Puzzles

JigShape:通过拼图任务评估视觉语言模型的视觉几何推理能力

Shawn Li, Wei Yang, Jike Zhong, Jiate Li, Jiawei Yang, You Qin, Ryan Rossi, Franck Dernoncourt, Roger Zimmermann, Yue Wang, Zhengzhong Tu, Vicente Ordonez, Mohit Bansal, Yue Zhao

机构 * University of Southern California(南加州大学) National University of Singapore(新加坡国立大学) Adobe Research(奥多比研究院) Texas A&M University(德克萨斯农工大学) Rice University(莱斯大学) The University of North Carolina at Chapel Hill(北卡罗来纳大学教堂山分校)

AI总结 本研究提出JigShape拼图基准,发现零样本VLM大多缺乏几何推理能力,所有模型在大尺寸拼图上均出现性能崩塌,将可扩展几何推理确立为VLM的开放性挑战。

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.15219 2026-08-05 cs.AI cs.IT math.IT 版本更新

NOVA: Fundamental Limits of Knowledge Discovery Through AI

NOVA:通过人工智能进行知识发现的基本限制

Salman Avestimehr, Ken Duffy, Muriel Médard

机构 * University of Southern California(南加州大学) Northeastern University(东北大学) Massachusetts Institute of Technology(麻省理工学院)

AI总结 本文提出NOVA框架,将“生成-验证-积累-再训练”循环建模为知识空间上的自适应采样过程,识别了知识覆盖有限域的条件及失败模式,并证明了发现成本与Zipf定律相关的标度律。

Comments Added an explicit recursive retraining model showing how accepted outputs reshape future generation. New results characterize when repeated retraining suppresses undiscovered artifacts and when mixing updates with a fixed base distribution preserves exposure. Corrected the Zipf discovery-cost proof and expanded the analysis. Main results and implications remain unchanged

详情

展开后加载摘要…

URL PDF HTML 收藏