CoV: Chain-of-View Prompting for Spatial Reasoning

TL;DR

This work tackles embodied question answering in 3D environments where fixed input viewpoints impede spatial reasoning. It introduces Chain-of-View prompting (CoV), a training-free, two-stage framework that first selects question-relevant anchor views and then iteratively adjusts the viewpoint using SE(3) camera transforms to gather discriminative observations. The approach yields substantial gains across OpenEQA, ScanQA, and SQA3D without additional training, including an average OpenEQA improvement of $+11.56\%$ and a maximum of $+13.62\%$ on Qwen3-VL-Flash, with further gains from test-time scaling. CoV demonstrates strong model-agnostic applicability and enhanced interpretability via multi-step, grounded reasoning, offering a practical pathway to boost spatial reasoning in 3D embodied QA for robotics and interactive agents.

Abstract

Embodied question answering (EQA) in 3D environments often requires collecting context that is distributed across multiple viewpoints and partially occluded. However, most recent vision--language models (VLMs) are constrained to a fixed and finite set of input views, which limits their ability to acquire question-relevant context at inference time and hinders complex spatial reasoning. We propose Chain-of-View (CoV) prompting, a training-free, test-time reasoning framework that transforms a VLM into an active viewpoint reasoner through a coarse-to-fine exploration process. CoV first employs a View Selection agent to filter redundant frames and identify question-aligned anchor views. It then performs fine-grained view adjustment by interleaving iterative reasoning with discrete camera actions, obtaining new observations from the underlying 3D scene representation until sufficient context is gathered or a step budget is reached. We evaluate CoV on OpenEQA across four mainstream VLMs and obtain an average +11.56\% improvement in LLM-Match, with a maximum gain of +13.62\% on Qwen3-VL-Flash. CoV further exhibits test-time scaling: increasing the minimum action budget yields an additional +2.51\% average improvement, peaking at +3.73\% on Gemini-2.5-Flash. On ScanQA and SQA3D, CoV delivers strong performance (e.g., 116 CIDEr / 31.9 EM@1 on ScanQA and 51.1 EM@1 on SQA3D). Overall, these results suggest that question-aligned view selection coupled with open-view search is an effective, model-agnostic strategy for improving spatial reasoning in 3D EQA without additional training.

Figures (13)

• Figure 1: The Chain-of-View prompting framework. Given a spatial query and its corresponding 3D scene, CoV facilitates a coarse-to-fine active reasoning process to derive the answer. (Bottom)$v_1$ to $v_4$ denote the task-relevant viewpoints strategically selected from the initial candidate view set by our View Selection Agent. (Left and Right) The interleaved action-reasoning chain demonstrates how the agent dynamically adjusts its perspective (e.g., rotation and movement) to gather discriminative visual evidence and resolve spatial ambiguities. (Center) The visualized camera frustums depict the autonomous exploration trajectory, where the agent bridges the gap between fragmented local views and global spatial context to reach a grounded conclusion.
• Figure 2: Video VLM vs. CoV. Unlike prior approaches (top) that rely on fixed-frame video inputs and answer from a limited temporal window, our chain-of-view framework (bottom) explores an open-ended view space constructed from a 3D scene. CoV dynamically selects informative viewpoints and performs step-by-step reasoning during inference, enabling more complete and grounded answers without additional training.
• Figure 3: Action-reasoning chain of the CoV agent. The CoV agent executes an iterative action–reasoning chain. For the question “What should I do to cool down?”, the agent first selects view 6 from the input images as an anchor. It then adjusts the viewpoint at each reasoning step to acquire new observations. Once the agent determines that sufficient information has been obtained, it outputs the answer “turn on the air conditioner.”
• Figure 4: Visualization of CoV reasoning results. Our method selects informative views and produces coherent multi-step answers grounded in the spatial context.
• Figure 5: Action Step Analysis. Distribution of action steps for Qwen3-VL-Flash bai2025qwen3vltechnicalreport on the OpenEQA openeqa dataset.