VisuoThink: Empowering LVLM Reasoning with Multimodal Tree Search

TL;DR

VisuoThink introduces a multimodal tree-search framework that enables slow, vision-text interleaved reasoning for LVLMs. By pairing vision-text interleaved thinking with predictive rollout search, it achieves inference-time scaling without fine-tuning, delivering state-of-the-art performance on geometry and spatial-reasoning benchmarks. The approach combines incremental visual construction with algebraic computation to reduce error propagation, and shows robust gains across geometry and spatial tasks with ablations clarifying the contributions of expansion, rollout, and selection. While computationally intensive, the method offers a principled path toward deeper multimodal reasoning in real-world tasks like navigation and tiling, with plans for reproducibility and ethical considerations.

Abstract

Recent advancements in Large Vision-Language Models have showcased remarkable capabilities. However, they often falter when confronted with complex reasoning tasks that humans typically address through visual aids and deliberate, step-by-step thinking. While existing methods have explored text-based slow thinking or rudimentary visual assistance, they fall short of capturing the intricate, interleaved nature of human visual-verbal reasoning processes. To overcome these limitations and inspired by the mechanisms of slow thinking in human cognition, we introduce VisuoThink, a novel framework that seamlessly integrates visuospatial and linguistic domains. VisuoThink facilitates multimodal slow thinking by enabling progressive visual-textual reasoning and incorporates test-time scaling through look-ahead tree search. Extensive experiments demonstrate that VisuoThink significantly enhances reasoning capabilities via inference-time scaling, even without fine-tuning, achieving state-of-the-art performance in tasks involving geometry and spatial reasoning.

Figures (5)

• Figure 1: Illustration of Input-Output Prompting, CoT, Vision-aided Thought and our VisuoThink. Vision-aided Thought often relies on reasoning with one-step or unreliable multi-step visual cues (generated by LVLMs). While VisuoThink addresses this gap through tool-augmented visual hints, coupled with a predictive-rollout search mechanism to systematically optimize reasoning capability.
• Figure 2: The illustration of our VisuoThink framework with three stages: (1) vision-text interleaved expansion: generates candidate paths through vision-text interleaved thinking; (2) rollout simulation: sample candidate reasoning nodes and then perform look-ahead search to better evaluate the value of current states; (3) selection: selects the most promising path via self-voting with results or states from rollout.
• Figure 3: The illustration of spatial reasoning tasks derived from VoTWu2024MindsEO, including Visual Navigation and Visual Tiling. LVLM is required to execute a sequence of actions to complete certain goals. Our experimental setting makes them much more challenging and closer to real-environment deployment.
• Figure 4: (LEFT) The trend of Pass@1 rate on Visual Navigation as the number of reasoning steps increases. (right) The relationship between the Accuracy@1 on geometry problems (Geomverse) and tree width for rollout search. We observe that LVLMs significantly benefit from longer reasoning chains, although the effect plateaus rapidly beyond a certain threshold of reasoning steps. The relationship between performance and tree width exhibits a more complex pattern, demonstrating an inverted U-shaped trend with both GPT-4o and Claude-3.5-Sonnet.
• Figure 5: The performance gain (+%) on tasks through predictive rollout search. The performance gain is calculated via the performance gap between VisuoThink (w/o rollout search) and VisuoThink.