Vision-and-language navigation (VLN) requires agents to keep track of what they have seen while following instructions. According to a new arXiv preprint, however, growing observation histories create two problems: redundancy and underuse. Retaining more information does not necessarily mean the agent will use it well.
To address this, the authors introduce ReInsBench, a diagnostic benchmark designed to evaluate how well VLN agents retain and apply historical evidence. They also propose ReInsVLN, a training-free method that revisits past observations at inference time. The work suggests that better use of history, rather than more training, can improve navigation performance.
Because this is a single preprint with a short abstract, the article does not compare findings with other sources. The specific benchmark tasks and experimental results are not detailed in the abstract.