Large vision-language models (LVLMs) have shown strong potential for embodied vision-language navigation (VLN), yet their performance degrades on long-horizo…
机构:北理工
来源:arXiv 2610.18478 | AI4Papers 论文推荐平台