DART-VLN: Test-Time Memory Decay and Anti-Loop Regularization for Discrete Vision-Language Navigation

DART-VLN is a training-free inference-time framework for discrete vision-language navigation. It addresses stale or redundant historical evidence during memory retrieval and inefficient local backtracking during action selection. Test-Time Memory Decay reweights memory slots without changing their stored content, while Anti-Loop Regularization penalizes immediate reversals in the next-hop decision. According to the abstract, experiments on R2R and REVERIE show that memory decay preserves or improves performance and reduces runtime. Combining both components produces shorter trajectories, less local backtracking, and the best quality-efficiency balance among evaluated GridMM variants, without changing the navigation backbone or adding learnable parameters.