Detached Skip-Links and $R$-Probe: Decoupling Feature Aggregation from Gradient Propagation for MLLM OCR
分离跳跃链接与$R$-探针:解耦特征聚合与梯度传播用于MLLM OCR
机构 * State Key Laboratory for Multimedia Information Processing, School of Computer Science, PKU-Anker LLM Lab, Beijing Key Laboratory of Software and Hardware Cooperative Artificial Intelligence Systems, Peking University, Beijing, China(多媒体信息处理国家重点实验室,计算机科学学院,PKU-Anker LLM实验室,软件与硬件协同人工智能系统北京重点实验室,北京大学,北京,中国) ; Tsinghua University, Beijing, China(清华大学,北京,中国) ; Baidu Inc, Beijing, China(百度公司,北京,中国)
专题命中 多模态训练与对齐 :MLLM(title,title_cn);multimodal(abstract);分类 cs.CV、cs.AI
AI总结 针对多模态大语言模型在OCR任务中因梯度干扰导致细粒度视觉信息丢失的问题,提出分离跳跃链接(Detached Skip-Links)以解耦前向特征聚合与反向梯度传播,并引入$R$-探针($R$-Probe)诊断视觉令牌的可重构性,从而提升OCR及通用多模态任务性能。
Comments Accepted by ICML 2026. Ziye Yuan and Ruchang Yao contributed equally to this work (co-first authors, listed in random order)