Aux-Think: Exploring Reasoning Strategies for Data-Efficient Vision-Language Navigation

TL;DR

This work systematically evaluates reasoning strategies for Vision-Language Navigation in continuous environments and uncovers a Test-time Reasoning Collapse when explicit CoT is used during testing. It introduces Aux-Think, a training framework that uses Chain-of-Thought as auxiliary supervision while performing No-Think at test time, and builds R2R-CoT-320k to support reasoning-based training. Through extensive experiments on R2R-CE and RxR-CE, Aux-Think demonstrates strong data efficiency and competitive performance with significantly less training data. The combination of reasoning-guided auxiliary tasks and receding-horizon action planning yields robust long-horizon navigation with improved generalization.

Abstract

Vision-Language Navigation (VLN) is a critical task for developing embodied agents that can follow natural language instructions to navigate in complex real-world environments. Recent advances in VLN by large pretrained models have significantly improved generalization and instruction grounding compared to traditional approaches. However, the role of reasoning strategies in navigation-an action-centric, long-horizon task-remains underexplored, despite Chain-of-Thought (CoT) reasoning's demonstrated success in static tasks like visual question answering. To address this gap, we conduct the first systematic evaluation of reasoning strategies for VLN, including No-Think (direct action prediction), Pre-Think (reason before action), and Post-Think (reason after action). Surprisingly, our findings reveal the Inference-time Reasoning Collapse issue, where inference-time reasoning degrades navigation accuracy, highlighting the challenges of integrating reasoning into VLN. Based on this insight, we propose Aux-Think, a framework that trains models to internalize structured reasoning patterns through CoT supervision, while inferring action directly without reasoning in online prediction. To support this framework, we release R2R-CoT-320k, the first Chain-of-Thought annotated dataset for VLN. Extensive experiments show that Aux-Think reduces training effort greatly and achieves the best performance under the same data scale.

Figures (7)

• Figure 1: Aux-Think outperforms alternative reasoning approaches in navigation tasks (a) and achieves a favorable trade-off between data usage and success rate (b).
• Figure 2: Illustration of Aux-Think and other reasoning strategies. Unlike No-Think, Pre-Think, and Post-Think, our Aux-Think introduces auxiliary CoT- and instruction-based reasoning during training while maintaining efficient action planning at testing.
• Figure 3: Illustration of CoT and Action Prediction Results Using Different Reasoning Strategies. Pre-Think generates incorrect actions (e.g., “move forward 75cm”) due to flawed CoT reasoning, such as “we have not crossed the room,” leading to significant trajectory deviation. Post-Think, which builds on Pre-Think’s output, inherits similar reasoning errors (e.g., “no obvious door or archway”) and makes the same wrong prediction. In contrast, Aux-Think correctly predicts “turn right 15 degrees” and follows a trajectory aligned with the ground truth. While Aux-Think does not rely on CoT during testing, it can optionally produce CoT via prompt switching—yet its action prediction remains accurate even when the generated CoT is of moderate quality. This highlights Aux-Think’s robustness to imperfect reasoning and its superior reliability in action prediction.
• Figure 4: The annotation pipeline of our R2R-CoT-320k dataset.
• Figure 5: Comparision of success rate on Pre-Think, Post-Think, and No-Think.