PaperlayerView exact revision on arXiv

Source-linked AI summary

VisuoThink: Empowering LVLM Reasoning with Multimodal Tree Search

Yikun Wang, Siyin Wang, Qinyuan Cheng, Zhaoye Fei, Liang Ding, Qipeng Guo, Dacheng Tao, Xipeng Qiu

arXiv:2504.09130v1cs.CL

TL;DR

Complex geometry and spatial reasoning often call for visual aids and deliberate, step-by-step thought, while existing approaches leave visual reasoning static or rudimentary. VisuoThink interleaves tool-generated visual hints with verbal reasoning and look-ahead tree search. It reports gains across geometry and spatial reasoning benchmarks without fine-tuning, including a Geomeverse accuracy@1 of 48.5%.

  • Problem

    LVLMs struggle with complex geometry and spatial reasoning, while existing slow-thinking or visual-aid approaches do not capture multi-step interleaved visual-verbal reasoning.

    Source: §1 · pp. 1–2

  • Method

    VisuoThink combines step-by-step vision-text interleaving with predictive rollout tree search to explore candidate reasoning paths during inference.

    Source: §1 · p. 2

  • Results

    VisuoThink reports gains across geometry and spatial reasoning benchmarks without fine-tuning; on Geomeverse, accuracy@1 reaches 48.5%, up to 21.8% over the state-of-the-art baseline.

    Source: §1 · p. 2, §7 · p. 9

  • Takeaways & Limitations

    The reported improvements indicate that inference-time scaling with visual-textual reasoning can enhance LVLM performance on geometry and spatial reasoning tasks without fine-tuning.

    Source: §7 · p. 9

    Predictive rollout search adds significant computational overhead, potentially making the approach impractical for real-time applications.

    Source: Limitations · p. 9

Abstract

from arXiv · show

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.

1 Introduction

VisuoThink addresses the gap between text-only deliberation and limited visual assistance by interleaving visual and verbal reasoning with search. The paper reports broad gains, particularly in geometry and spatial reasoning.

1 Introduction

  • Existing slow-thinking approaches use visual information as static input, while vision-aided methods often lack multi-step visual-textual reasoning and search.Humans commonly use visual aids for complex geometry and spatial reasoning, including iteratively sketching auxiliary lines.Source: §1 · pp. 1–2
  • On Geomeverse, VisuoThink reaches accuracy@1 48.5%, with improvement up to 21.8% over the state-of-the-art baseline.The authors also report significant gains across various reasoning tasks, particularly geometry and spatial reasoning.Source: §1 · p. 2
  • VisuoThink introduces multimodal tree search, dynamically integrating visual and verbal reasoning paths throughout problem-solving.Its predictive rollout mechanism explores reasoning paths by predicting future states, extending test-time scaling to visual reasoning.Source: §1 · p. 2

2 Related Work

Related work develops text-centric deliberation, visual assistance, and test-time search, but the cited approaches leave visual information static or provide limited visual reasoning and search.

2.1 Text-centric Reasoning in LVLMs

  • Text-centric LVLM reasoning methods adapt stage-wise reasoning, curriculum learning, tree-search data generation, and distillation, but use images as static input and reason through text.The cited examples include Xu et al. (2024), Thawakar et al. (2025), Yao et al. (2024), and Du et al. (2025).Source: §2.1 · p. 2

2.2 Vision-aided Reasoning

  • Vision-aided methods use transformed visual information, perception tools, or intermediate representations, yet often lack multi-step visual cues and search over reasoning paths.Visual Sketchpad draws sketches, while VoT uses emoji-like representations.Source: §2.2 · pp. 2–3

2.3 Test-time Scaling with Tree Search

  • Text-model test-time scaling explores methods such as BoN, guided beam search, and MCTS, while LVLM work remains limited to basic text-only search such as AtomThink beam search.VisuoThink extends test-time scaling with vision-text interleaved thinking and look-ahead search.Source: §2.3 · p. 3

3 VisuoThink

VisuoThink combines dynamic visual-textual reasoning with predictive rollout search. Its process constructs visual hints through tool use, simulates candidate paths, and selects a promising path using model self-voting.

3 VisuoThink

  • VisuoThink dynamically integrates visual and textual information during inference, with predictive rollout search as its mechanism for thinking ahead.The framework presents this as multimodal slow thinking.Source: §3 · p. 3

3.1 Vision-Text Interleaved Thinking

  • Vision-text interleaved reasoning cycles through Thought, Action, and Observation: plan a visual hint, call a tool to create or modify it, then use the visual feedback.Tool invocations construct visual hints step by step and adapt to different visual reasoning tasks.Source: §3.1 · p. 3
  • Step-by-step visual guidance integrates with search, unlike approaches that generate all visual aids at once.The authors describe this tool-based design as enabling flexible adaptation across visual reasoning tasks.Source: §3.1 · p. 3

3.2 Predictive Rollout Search

  • Predictive rollout search expands multiple candidate paths, simulates each candidate to a final result, then uses self-voting to select a promising path.The rollout predicts full-path outcomes rather than relying only on immediate state evaluation; the LVLM heuristic considers the task, historical nodes, and simulated paths.Source: §3.2 · p. 4

4 Solving Geometry with VisuoThink

VisuoThink combines progressive visual construction with algebraic computation and search-based reasoning for geometry problems. Its results improve over CoT and Visual Sketchpad, with further gains from predictive rollout search.

4 Solving Geometry with VisuoThink

  • Geometry solving proceeds from auxiliary-line construction to translating geometric relationships into equations solved through Python execution.Phase I constructs constraint-defined lines and ends with AUX-END; Phase II uses tools such as an equation solver for numerical resolution.Source: §4 · pp. 4–5
  • The model produces visual-textual reasoning steps leading to a result, conditioned on the problem and auxiliary lines and bounded by a maximum step limit.The framework validates executable steps with coordinate-based tools such as matplotlib and an equation solver.Source: §4 · p. 4
  • Progressively constructed graphical representations let search refine auxiliary-line generation, while computational tools target exact numerical solutions and mitigate reasoning errors.The method uses common Python libraries such as matplotlib for rendering and equation solvers for numerical resolution.Source: §4 · pp. 4–5

4.1 Empirical Results

  • On Geomverse-109, VisuoThink without rollout search outperforms CoT by 17.1% and Visual Sketchpad by 16.7% on average across evaluated models; rollout search adds 4.1%.The Visual Sketchpad equation solver alone increases average performance by 3.3%.Source: §4.1 · pp. 5–6, Table 1 · p. 5

5 Spatial Reasoning with VisuoThink

VisuoThink applies tool-mediated visual feedback to challenging Visual Navigation and Visual Tiling tasks. It achieves the highest reported accuracy across the spatial tasks, with gains associated with executor feedback and multi-step visual reasoning.

5 Spatial Reasoning with VisuoThink

  • Visual Navigation and Visual Tiling adapt VoT benchmarks into more challenging tasks, and the experiments vary Navigation map complexity while evaluating level-2 Tiling problems.The evaluation compares against CoT and VoT, including VoT with executor interactions, whose visual hints are generated by the model rather than the executor.Source: §5 · p. 6, §5.1 · p. 6
  • An executor carries out each action and returns textual feedback plus optional visuospatial state cues: the agent’s position for Navigation or rectangle occupation for Tiling.The experiments associate executor feedback and multi-step visual reasoning with improved performance; VoT with textual feedback gains an average of 8.1% on Visual Tiling, less than VisuoThink without rollout search.Source: §5 · p. 6, §5.1 · p. 7

5.1 Empirical Results

  • On Visual Navigation level 3, GPT-4o with VisuoThink achieves 93.8% accuracy, versus 62.5% for VoT with an executor and 18.8% for CoT.The reported improvement pattern extends across GPT-4o, Qwen2-VL-72B-Instruct, and Claude-3.5-sonnet; VisuoThink without rollout search improves Visual Tiling by an average of 34.7% across models.Source: §5.1 · p. 7

6 Discussion

The discussion examines how reasoning length, tree width, and rollout supervision relate to VisuoThink’s performance. The reported patterns favor more reasoning steps only up to a plateau, show non-monotonic effects of tree width, and vary with supervision strength.

6.1 Could Longer Reasoning Chains Assist LVLMs in Reasoning?

  • Increasing Navigation reasoning steps from 10 to 20 yields gains of 54.1% for GPT-4o and 48.4% for Claude-3.5-sonnet, but increasing from 20 to 40 yields only 6.5% and 2.1%, respectively.Completion rates rise with more reasoning steps and then converge, making further substantial improvements difficult.Source: §6.1 · pp. 7–8

6.2 Could Larger Tree Span Enhances VisuoThink’s Performance?

  • With tree depth held constant, geometry performance follows an inverted U as candidate child nodes increase, so widening the rollout tree does not continuously improve results.The authors attribute the later decline to errors in evaluating child nodes, which can confuse selection and reduce reasoning efficiency.Source: §6.2 · p. 8

6.3 Strong v.s. Weak Supervision in Predictive Rollout Search

  • Predictive rollouts produce larger performance gains in spatial reasoning than in geometry across open- and closed-source models, where spatial tasks provide stronger visual-state supervision.Geometry rollouts receive final numerical results, while spatial rollouts can receive visual states such as the agent’s final position and destination.Source: §6.3 · p. 8, Figure 5 · p. 8

7 Conclusion

The conclusion presents VisuoThink as a multimodal tree-search framework for LVLM reasoning. It reports gains on geometry and spatial reasoning without fine-tuning and summarizes findings about tool interaction, search, and supervision.

7 Conclusion

  • VisuoThink combines dynamic visual-textual interleaving with predictive rollout search to enhance LVLM reasoning.The conclusion presents this as a multimodal tree-search framework.Source: §7 · pp. 8–9
  • The approach improves performance across geometry and spatial reasoning tasks without requiring model fine-tuning.The conclusion describes substantial gains on geometry and spatial reasoning benchmarks.Source: §7 · p. 9
  • The paper’s analyses highlight tool-interaction benefits, search-space optimization, and supervision strength as factors relevant to multimodal reasoning.The conclusion presents these findings as directions for advancing LVLM capabilities on complex reasoning tasks.Source: §7 · p. 9

Limitations

VisuoThink has limitations in computational cost, tool dependence, base-model dependence, and evaluation scope.

Limitations

  • Predictive rollout search adds significant computational overhead, potentially making VisuoThink impractical for real-time applications.The stated constraint is potential impracticality, not a claim that real-time use is impossible.Source: Limitations · p. 9
  • Stronger capability particularly relies on tool interactions, which may require more effort in some deployment environments.Source: Limitations · p. 9
  • VisuoThink’s effectiveness is constrained by the base VLM’s reasoning capabilities and cannot overcome fundamental model limitations.Source: Limitations · p. 9
  • The evaluation focuses primarily on geometric and spatial reasoning tasks.Source: Limitations · p. 9

Ethics and Reproducibility Statements

The authors state that the work raises no ethical issues and describe their reproducibility practices and plans.

Ethics

Reproducibility

A Performance Gain of VisuoThink Through Predictive Rollout Search

The appendix measures gains from predictive rollout search across geometry and spatial reasoning tasks, reporting larger examples on strongly supervised tasks than weakly supervised ones.

A Performance Gain of VisuoThink Through Predictive Rollout Search

B OKSpatial Reasoning Task Setting

The appendix specifies stricter spatial-task outputs than VoT, defines task formulations, reports experimental settings, and describes Geomverse-109 generation.

B OKSpatial Reasoning Task Setting

  • VisuoThink’s spatial formulation requires explicit movement directions and step counts, extending beyond VoT’s direction-only navigation and simplified tiling specifications.Visual Tiling also calls for variant types, block positions, and action types such as “fit” or “remove.”Source: B OKSpatial Reasoning Task Setting · p. 11

C Task Formulation of Spatial Reasoning Tasks

  • Visual Navigation requires executable direction-and-step-count actions on an obstacle-laden grid; Visual Tiling requires polyomino placement or removal actions in a masked rectangle.For tiling, action specifications include selected polyomino type, placement-block coordinates, and action type; variant type can be implicit in block positions.Source: C Task Formulation of Spatial… · p. 11

D Model and VisuoThink Hyperparameters

  • Experiments fixed reasoning-tree depth at 10 and sampled three child nodes; VisuoThink used temperature 0.8 for child sampling and temperature 0 for other invocations and voting.Voting used single-vote sampling to reduce model calls while achieving comparable performance.Source: D Model and VisuoThink Hyperparameters · p. 12

E Geomverse-109 Problem Generation Trajectory

  • Geomverse-109 was constructed from Geomverse problems using generated Python code and filtering for diagram quality, yielding a 90-problem dataset.The passage says 109 problems were randomly chosen before filtering.Source: E Geomverse-109 Problem Generation… · p. 12
Loading 2504.09130v1…