Computer Vision and Pattern Recognition · arXiv:2601.13029 v5 · submitted 19 Jan 2026
Think3D: Thinking with Space for Spatial Reasoning
The paper says
“These results show that active exploration in 3D space provides an effective and general paradigm for improving spatial reasoning in multimodal agents.”
from the abstract · picked out as the main claim by a classifier
Reviews
Reviews by people
No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
While Vision-Language Models (VLMs) excel at 2D visual understanding, they remain constrained by 2D-centric paradigm that severely limits genuine 3D spatial reasoning. To bridge this gap, we introduce Think3D, a novel framework that equips VLM agents with interactive, 3D chain-of-thought reasoning capabilities. By integrating a suite of 3D manipulation tools, Think3D transforms perception into active spatial exploration, mirroring human geometric reasoning. Think3D consistently improves proprietary models, including GPT-4.1 and Gemini 2.5 Pro, across BLINK Multi-view, MindCube-1K, and VSI-Bench-Tiny. We further propose Think3D-RL to teach smaller open-weight models how to manipulate 3D space effectively. Using only final-answer rewards, without process supervision or handcrafted exploration trajectories, Think3D-RL enables Qwen3-VL-4B to autonomously learn effective 3D exploration strategies. After training, the model exhibits tool-use patterns similar to those of stronger proprietary models, while shifting the effect of 3D tool use on MindCube-1K from a performance drop to a substantial improvement. These results show that active exploration in 3D space provides an effective and general paradigm for improving spatial reasoning in multimodal agents. Code, models, and data are available at https://github.com/zhangzaibin/spagent.
The paper's own sentences, arranged. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-09-23 · no language-model review
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.14 · states limitation 0.16 · human participants 0.04 · controlled comparison 0.49 · randomized assignment 0.02 · states quantitative result 0.12 · compares with prior results 0.77 · mentions code or data availability 0.99
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v4). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| While Vision-Language Models (VLMs) excel at 2D visual understanding, they remain constrained by 2D-centric paradigm that severely limits genuine 3D spatial reasoning. | background 0.57 | 0.09 | 0.04 | asserted 1.00 |
| To bridge this gap, we introduce Think3D, a novel framework that equips VLM agents with interactive, 3D chain-of-thought reasoning capabilities. | contribution 1.00 | 0.88 | 0.05 | asserted 0.98 |
| By integrating a suite of 3D manipulation tools, Think3D transforms perception into active spatial exploration, mirroring human geometric reasoning. | method 0.86 | 0.62 | 0.07 | asserted 1.00 |
| Think3D consistently improves proprietary models, including GPT-4.1 and Gemini 2.5 Pro, across BLINK Multi-view, MindCube-1K, and VSI-Bench-Tiny. | result 1.00 | 0.91 | 0.09 | asserted 1.00 |
| We further propose Think3D-RL to teach smaller open-weight models how to manipulate 3D space effectively. | contribution 0.91 | 0.88 | 0.04 | asserted 0.96 |
| Using only final-answer rewards, without process supervision or handcrafted exploration trajectories, Think3D-RL enables Qwen3-VL-4B to autonomously learn effective 3D exploration strategies. | method 0.87 | 0.94 | 0.05 | asserted 1.00 |
| After training, the model exhibits tool-use patterns similar to those of stronger proprietary models, while shifting the effect of 3D tool use on MindCube-1K from a performance drop to a substantial improvement. | result 1.00 | 0.94 | 0.10 | asserted 1.00 |
| These results show that active exploration in 3D space provides an effective and general paradigm for improving spatial reasoning in multimodal agents. | interpretation 1.00 | 0.85 | 0.90 | asserted 1.00 |
| Code, models, and data are available at https://github.com/zhangzaibin/spagent. | contribution 0.68 | 0.33 | 0.02 | asserted 0.87 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.98
What was examined, and what was not
Sources read
- arXiv listing abstract v4
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 6,993 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Cited by 9 Pith papers
Reviewed papers on Pith that cite this one. OpenAlex reports about 34 citations worldwide. The full citation record
- Think-Capture: Active 3D Geometry Acquisition for Spatial Reasoning in Vision-Language Models cs.CV · Oct 2026
- From Views to Worlds: Active Exploration over 3D Worlds for Vision-Language Models cs.CV · Oct 2026
- ReTAC: Residual Test-Time Agentic Computation for Video Spatial Reasoning cs.AI · Oct 2026
- Seek-and-View Reasoning for Multi-View Spatial Understanding cs.CV · Oct 2026
- S-Agent 2.0: Spatial Reasoning with a General Coding Agent cs.AI · Oct 2026
- DeSpatial: The devil of Agentic Spatial Reasoning Lies in 3D Itself cs.CV · Oct 2026
- SpatialASM: Internalizing Active Mentalization for Spatial Reasoning cs.CV · Oct 2026
- Distilling Routed 3D Privilege for Spatial Reasoning in Vision-Language Models cs.CV · Oct 2026
- Rethinking Multi-Image Re-Representation in Multi-Image Understanding cs.CV · Sep 2026
How the literature uses it
14 arXiv papers cite this in 40 sentences
Papers citing it · 16 sentences without a confident label
Uses its method or tool 2 papers
-
For VSI-Bench, we follow Think3D [ref] and uniformly sample seven frames from the full scene video to serve as model input.
Skill-3D: Evolving Scene-Aware Skills for Agentic 3D Spatial Reasoning · Experimental Setup · 0.98
-
For the post-trained AlloSpatial agents and other baselines, we follow the limited-observation setting of Think3D [ref] and uniformly sample 7 frames, testing performance under sparse visual inputs.
AlloSpatial: Agentic Harness Framework for Spatial Reasoning in Foundation Models · Experimental Setup · 0.93
Compares with it 4 papers
-
Each backbone is tested under four settings: w/o Tools, w/ Tools, Think3D [ref] , and Skill-3D.
Skill-3D: Evolving Scene-Aware Skills for Agentic 3D Spatial Reasoning · Experimental Setup · 0.92
-
We compare against direct inference with GPT-5-mini and Gemini-3-flash, spatial QA fine-tuned models [ref] , and test-time scaling methods including MindJourney [ref] and Think3D [ref] .
Reasmory: 3D Reconstruction as Explicit Memory for VLMs Spatial Reasoning · Experiments · 1.00
-
We compare 1) Proprietary models including GPT-5.2, Claude-4.6-Opus, and Gemini-3-Pro; 2) Open-source models including Qwen3-VL [ref] , InternVL3.5 [ref] , and Gemma-4 [ref] ; 3) Specialized spatial models including Spatial-MLLM [ref] , Cambrian-S [ref] , and Think3D [ref] .
AlloSpatial: Agentic Harness Framework for Spatial Reasoning in Foundation Models · Experimental Setup · 0.98
Cites it as support 1 paper
-
MindCube [ref] and related studies [ref] demonstrate that providing additional views alone yields limited gains; instead, they show that performance improves when models first construct an intermediate representation such as a cognitive map and then reason over it.
Dense Reward for Multi-View 3D Reasoning with Global Maps and Local Views · Introduction · 0.96
Cites it as background 10 papers
-
A parallel line introduces 3D-derived intermediates, such as depth maps, Gaussian splats, or 3D reconstructions [ref] , while others predict camera trajectories to mentally simulate unseen viewpoints [ref] .
How and What to Imagine? Visual Thinking in Unified Multimodal Models for Cross-View Spatial Reasoning · Related Work · 0.95
-
A parallel line of work injects 3D structure into VLMs via depth maps, point clouds, or features from explicit 3D reconstruction [ref] .
Thinking with Novel Views: A Systematic Analysis of Generative-Augmented Spatial Intelligence · Introduction · 0.99
-
Other approaches operate at test time by generating additional views [ref] , invoking 3D tools [ref] , or encoding geometric references for an MLLM [ref] .
Reasmory: 3D Reconstruction as Explicit Memory for VLMs Spatial Reasoning · Related Work · 0.93
What it cites · 26 sentences without a confident label
Uses the method or tool of 2 papers
-
* Think3D-RL Reinforcement Learning Module optimizes multi-step 3D exploration policy through tool calling, trained with Group Relative Policy Optimization (GRPO) [ref] .
DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models · Framework Overview · 0.99
-
Setting and Dataset Our reinforcement learning (RL) training framework is based on SWIFT [ref] .
SWIFT:A Scalable lightWeight Infrastructure for Fine-Tuning · Experiment Setup · 1.00
Compares with 5 papers
-
In addition, comparisons are made against specialized models fine-tuned on spatial reasoning datasets, including RoboBrain [ref] , Spatial-MLLM [ref] , VLM-3R [ref] , as well as REVPT [ref] , a tool-augmented fine-tuning method.
VLM-3R: Vision-Language Models Augmented with Instruction-Aligned 3D Reconstruction · Experiment Setup · 1.00
-
In addition, comparisons are made against specialized models fine-tuned on spatial reasoning datasets, including RoboBrain [ref] , Spatial-MLLM [ref] , VLM-3R [ref] , as well as REVPT [ref] , a tool-augmented fine-tuning method.
Reinforced Visual Perception with Tools · Experiment Setup · 1.00
-
In addition, comparisons are made against specialized models fine-tuned on spatial reasoning datasets, including RoboBrain [ref] , Spatial-MLLM [ref] , VLM-3R [ref] , as well as REVPT [ref] , a tool-augmented fine-tuning method.
Spatial-MLLM: Boosting MLLM Capabilities in Visual-based Spatial Intelligence · Experiment Setup · 1.00
Cites as support 1 paper
-
In particular, DeepEyes [ref] promotes “thinking with images”, enabling models to leverage internal visual reasoning capabilities without external tools and directly inspiring our design.
DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning · VLM tool calling · 0.98
Cites as background 24 papers
-
Reinforcement learning (RL) has become a central paradigm for tool-use and reasoning policies [ref] learning.
Reinforced Visual Perception with Tools · VLM tool calling · 0.83
-
Reinforcement learning (RL) has become a central paradigm for tool-use and reasoning policies [ref] learning.
OpenThinkIMG: Learning to Think with Images via Visual Tool Reinforcement Learning · VLM tool calling · 0.83
-
Reinforcement learning (RL) has become a central paradigm for tool-use and reasoning policies [ref] learning.
2505.19255 · VLM tool calling · 0.83
Challenges and discussion
Know something about this paper? Leave a comment.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment