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

Jul 16, 2026

Upload by Finn

Memory-based discrete vision-language navigation (VLN) agents must decide under partial observability, yet frozen backbones can still fail at inference time. DART-VLN introduces training-free test-time control to address two issues: stale evidence during memory readout and inefficient local backtracking during action selection. Test-Time Memory Decay reweights grid aggregation to suppress redundant history without changing stored content. Anti-Loop Regularization penalizes immediate reversals via next-hop scoring. Experiments on R2R and REVERIE show decay-only improves stability, while decay+anti-loop yields the best quality-efficiency trade-off with shorter trajectories, lower runtime, and higher navigation quality.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages7

Reading Time18 min

Introduction

Contributions

DART-VLN: Test-Time Memory Decay and Anti-Loop Regularization for Discrete Vision-Language Navigation Shaoheng Zhang 1 , Zhichen Li2 , and Jie Mei 1 ,∗ 1 School of Intelligence Science and Engineering 2 School of Computer Science and Technology Harbin Institute of Technology, Shenzhen Shenzhen, China [2023312309@stu.hit.edu.cn](2023312309@stu.hit.edu.cn) , [2023111963@stu.hit.edu.cn](2023111963@stu.hit.edu.cn) , [jmei@hit.edu.cn](jmei@hit.edu.cn) arXiv :2607 .0 1043v 1 [ cs .RO] 1 Jul 2026 Abstract—Memory-based discrete vision-language navigation (VLN) agents must act under partial observability, yet even strong frozen backbones remain vulnerable at test time. Two common failure modes are stale historical evidence at memory readout and inefficient local backtracking during action selection. We present DART-VLN, a training-free test-time control framework for discrete VLN. DART-VLN combines Test-Time Memory Decay, a read-side memory reweighting rule that suppresses stale and redundant evidence without rewriting stored content, with Anti-Loop Regularization, a lightweight nexthop penalty that discourages immediate reversals during actionselection. The framework introduces no new learnable parameters and leaves the learned backbone unchanged. Experiments on R2R and REVERIE show a consistent pattern: decay-only provides stable read-side gains, while decay+anti-loop achieves the best overall quality-efficiency trade-off, yielding shorter trajectories, lower runtime, and improved navigation performance in key settings. Behavioral analysis further confirms that antiloop regularization reduces local backtracking and improves path efficiency under frozen backbones. Overall, the results show that modest test-time control can make memory-based discrete VLN more reliable and efficient without retraining. Index Terms—Vision-Language Navigation, Discrete Navigation, Test-Time Control, Memory Decay, Anti-Loop Regularization I. INTRODUCTION Language-guided embodied navigation requires reliable sequential decision-making under partial observability. Vision-language navigation (VLN) provides a standard testbed for this problem by asking an embodied agent to move through an environment using language and visual observations [1], [2] . Among existing formulations, discrete VLN is especially attractive because it operates over explicit viewpoint graphs and supports controllable step-wise decision-making [3]–[5] . Recent progress in VLN has come from stronger pretrained navigators, explicit memory or map representations, broader training resources, and more sophisticated recovery or planning mechanisms [3], [6]–[9] . Together, these advances have substantially improved long-horizon reasoning and benchmark performance. Yet a practical gap remains at inference time: even strong memory-based discrete VLN * Corresponding author. Code will be released at [https://github.com/Japluto/](https://github.com/Japluto/)[ ](https://github.com/Japluto/)DART-VLN. backbones can still behave unreliably under frozen parameters. In many cases, further gains rely on retraining, architectural redesign, or heavier planning modules, which are less appealing when the goal is to strengthen an already competitive navigator with minimal intervention. In this setting, two recurring failure modes become particularly important. The first appears at memory readout. Explicit navigation memory helps agents reason over longer trajectories by storing previously visited viewpoints, visual features, or map-level context [3]–[5], [7], [10]–[12] . Yet as navigation proceeds, stale or repeatedly observed evidence may remain active after its usefulness has faded, making memory aggregation noisier at decision time. The second appears at action selection. Even with a strong frozen backbone, agents still exhibit inefficient local behaviors such as immediate reversals and short loops [13]–[15] . These behaviors do not always destroy the final outcome, but they l

FAQs

What problem does DART-VLN address in discrete vision-language navigation at test time?

DART-VLN targets inference-time unreliability caused by stale evidence during memory readout and inefficient local backtracking such as immediate reversals or short loops during action selection.

How does Test-Time Memory Decay work in the DART-VLN framework?

Test-Time Memory Decay reweights the read-side memory during grid aggregation to suppress stale and redundant evidence, while not rewriting or altering stored memory content.

How does Anti-Loop Regularization improve action selection?

Anti-Loop Regularization adds a lightweight next-hop penalty to discourage immediate reversals, reducing local backtracking and improving path efficiency under a frozen backbone.

Jul 16, 2026

Upload by Finn