Source-linked AI summary
Revisiting the Necessity of Lengthy Chain-of-Thought in Vision-centric Reasoning Generalization
Yifan Du, Kun Zhou, Yingqian Min, Yue Ling, Wayne Xin Zhao, Youbin Wu
TL;DR
The paper investigates which CoT designs support generalizable visual reasoning in VLMs, a question left open despite the use of long and visual CoT supervision. It compares language, grounding, and visual CoT in a controlled maze benchmark under SFT-then-RL training, finding that concise, minimally grounded traces generalize best while visual and longer traces mainly accelerate convergence.
Problem
It remains unclear which CoT designs contribute to generalizable visual reasoning, despite widespread use of long and visual CoT supervision.
Source: §1 · p. 2Method
The study compares language, grounding, and visual CoT using a controlled maze benchmark and an SFT-then-RL training paradigm.
Source: §1 · p. 2, §6 · p. 10Results
Minimal-grounding CoT generalizes best across maze sizes, short CoT often surpasses longer traces, and visual or longer CoT mainly accelerates convergence without substantially raising the final performance ceiling.
Source: §1 · p. 2, §6 · p. 10Takeaways & Limitations
The findings support concise, well-grounded supervision as a practical guide for constructing more generalizable visual-reasoning SFT datasets.
Source: §1 · p. 2, §6 · p. 10The analysis is centered on mazes and VLMs, with extension to richer task families beyond them left for future work.
Source: §1 · p. 2, §6 · p. 10
Abstract
from arXiv · showhide
We study how different Chain-of-Thought (CoT) designs affect the acquisition of the generalizable visual reasoning ability in vision-language models (VLMs). While CoT data, especially long or visual CoT such as "think with image", has been widely used to supervise intermediate reasoning, it remains unclear why specific CoT designs help and which ones truly support generalizable reasoning. To systematically evaluate this, we focus on a controlled maze-solving benchmark where reasoning rules are fully visual, difficulty can be tuned by grid size, and all the intermediate steps can be automatically generated. Using Qwen2.5-VL-7B under a standard SFT-then-RL pipeline, we compare three representative CoT formats: Language CoT, Grounding CoT (with spatial coordinate trajectories), and Visual CoT (with image manipulations). Our experiments reveal that visual and longer CoT mainly accelerate convergence but do not lift the final performance ceiling; concise CoT containing only essential grounding steps outperforms longer traces; and, strikingly, CoT retaining only the minimal grounding results generalizes best across different maze sizes. We further validate these insights on other vision-centric tasks. These findings highlight a "short is long" effect and provide practical guidance for constructing more generalizable SFT datasets for visual reasoning.
1 Introduction
The paper asks which CoT designs support generalizable visual reasoning and compares them in a controlled maze benchmark using an SFT-then-RL pipeline. Its findings favor concise, minimally grounded traces over verbose supervision, while visual and longer CoT mainly accelerate convergence.
1 Introduction
- The study compares Language, Grounding, and Visual CoT, which externalize intermediate reasoning in linguistic, spatial, and visual forms.Visual CoT manipulates the image, while Grounding CoT expresses a coordinate trajectory.Source: §1 · p. 2
- Mazes provide a controlled, contamination-free testbed with fully visual rules, grid-size-tunable difficulty, and automatically generated solutions and intermediate steps.The experiments use Qwen2.5-VL-7B in an SFT-then-RL generalization paradigm.Source: §1 · p. 2
- Visual CoT and longer traces can speed convergence, but final performance remains similar to or only marginally above shorter-CoT baselines.The paper characterizes this as faster optimization without a substantially higher performance ceiling.Source: §1 · p. 2
- Minimal-grounding CoT variants, such as sparse coordinate paths or final trajectories, generalize best across maze sizes.The paper reports that concise CoT with essential grounding also achieves higher and more stable performance than longer step-by-step traces.Source: §1 · p. 2
- The results support a “short is long” effect and offer guidance for constructing more generalizable visual-reasoning SFT datasets.The paper also reports validation on visual search and visual puzzle tasks.Source: §1 · p. 2
2 Preliminary
The paper treats CoT as a format for intermediate reasoning and distinguishes text-only, spatially grounded, and image-manipulating strategies.
2 Preliminary
- A VLM receives a textual query and image, generates a textual answer, and represents reasoning as intermediate states whose format is specified by the CoT strategy.The paper studies three strategy types.Source: §2 · p. 3
2.1 Language CoT
- Language CoT expresses the reasoning process purely through textual tokens, without manipulating visual information.It uses explicit reasoning steps conditioned on the query, image, and preceding steps.Source: §2.1 · p. 3
2.2 Grounding CoT
- Grounding CoT links linguistic references to image evidence using spatial descriptors that can represent points, lines, or regions.The model learns to connect each reasoning step with the spatial evidence it refers to.Source: §2.2 · p. 4
2.3 Visual CoT
- Visual CoT selects grounded targets, applies operations to the current image, and interleaves textual reasoning with updated visual contexts.The described operations include marking points, drawing lines, and cropping regions.Source: §2.3 · p. 4
3 Experimental Setup
The setup uses maze navigation as a controllable visual-reasoning testbed and compares three synthesized CoT formats within an SFT-then-RL training pipeline. The implementation trains Qwen2.5-VL-7B through convergence while keeping the vision encoder frozen.
3 Experimental Setup
- Maze navigation provides a testbed with tunable complexity, limited interference from pretrained priors, and primarily spatial reasoning requirements.Maze complexity can be adjusted by changing maze size or path length.Source: §3 · p. 4
3.1 Maze Data Construction
- Language CoT expresses paths as direction sequences, Grounding CoT records each visited grid by image coordinates, and Visual CoT draws the partial path on successive images.G-CoT-least instead represents the path sequence itself without additional textual explanation or absolute coordinates.Source: §3.1 · p. 5
3.2 Training Strategy
- The training pipeline first SFT-tunes separate CoT-style models, then applies GRPO reinforcement learning using 20K additional maze samples and accuracy and format rewards.Each CoT type has 8K SFT trajectories; the accuracy reward checks whether the path connects start and end without crossing walls.Source: §3.2 · pp. 5–6
3.3 Implementation Details
- Qwen2.5-VL-7B is trained for up to 1000 RL steps to reach convergence, with the vision encoder frozen and only LLM parameters updated.SFT uses three epochs, a learning rate of 1 × 10^-5, and batch size 64.Source: §3.3 · p. 6
4 Experimental Analysis
The experiments compare how CoT formats affect visual reasoning and generalization, first on controlled mazes and then on other vision-centric tasks. Least Grounding CoT generalizes best across maze sizes, while shorter grounding traces also perform strongly on additional tasks.
4.1 Experimental Framework
- The analysis asks what CoT strategies offer on unseen tasks, what capability enables them in vision-centric tasks, and how well they generalize, testing mazes before other tasks.The framework examines these questions first in controlled maze navigation and then on more realistic vision-centric tasks.Source: §4.1 · p. 6
4.2.1 Visual CoT Improves Efficiency, Not Efficacy
- SFT initialization is essential for stable, progressive RL: scratch-trained models tend to collapse, while SFT-initialized models reach 100% accuracy on training mazes.The authors attribute this to a better-shaped policy space and mitigation of exploration and sparse-reward problems.Source: §4.2.1 · p. 7
- V-CoT converges in roughly half the training steps required by L-CoT, but its faster convergence does not yield higher final accuracy than G-CoT or L-CoT.The reported interpretation is that V-CoT accelerates optimization rather than expanding ultimate maze-reasoning capacity.Source: §4.2.1 · p. 7
4.2.3 Least CoT Achieves Better Generalization
- In single-scale (cold-started on 6 × 6 mazes) and cross-scale (4 × 4 to 6 × 6) settings, G-CoT-least maintains high success on unseen sizes; V-CoT saturates after about 800 RL steps and remains inferior on training and test mazes.Source: Figure 4 · p. 8, Figure 5 · p. 8, §4.2.3 · p. 9
4.2.2 Short CoT Surpasses Longer Ones
- G-CoT-least starts higher and converges faster than explicit G-CoT, surpasses V-CoT in training efficiency, and reaches 100% training accuracy without explicit coordinates.The authors suggest that aligned grounding lets the model use latent spatial representations without an externally specified coordinate system.Source: §4.2.2 · p. 8
4.3 Experimental Results on Other Tasks
- Across VQA benchmarks and subtasks, G-CoT-least achieves the best performance; on FrozenLake and Jigsaw, it improves results, with Jigsaw accuracy rising from 0% to over 70%.The VQA evaluation includes V* and HR-Bench; the visual games use 4 × 4 FrozenLake grids and 3 × 3 Jigsaw puzzles.Source: Table 1 · p. 9, §4.3 · pp. 9–10
5 Related Work
Related work distinguishes language-dominant reasoning from vision-centric reasoning grounded in perceptual exploration, and reviews CoT and RL approaches for improving reasoning in VLMs.
Vision-centric Reasoning
- Vision-centric tasks such as mazes require active visual exploration and manipulation, unlike tasks whose reasoning can proceed linguistically after visual understanding.The paper focuses on the vision-centric category and uses maze navigation as a controlled testbed.Source: §5 · p. 10
Chain-of-thought Reasoning
- Prior work uses CoT to decompose complex problems and explores longer traces and grounding- or visual-based CoT variants for VLMs.This work compares these variants in a controlled setting to study their mechanisms.Source: §5 · p. 10
Reinforcement Learning in Vision Language Models
- Prior RL studies report gains on vision-centric tasks, where induced CoT trajectories tend to be short; this work attributes effective brief reasoning to reinforced grounding.The paper states that once grounding is sufficiently reinforced, models can reason effectively with very brief CoT traces.Source: §5 · p. 10
6 Conclusion
The paper compares three CoT formats in a controlled maze benchmark using Qwen2.5-VL-7B with SFT followed by RL. It finds that concise, minimally grounded traces generalize best, while longer or visual traces can accelerate convergence without substantially raising the final performance ceiling.
6 Conclusion
- Minimal grounding information, such as sparse coordinate paths, generalizes best across maze sizes and related visual reasoning tasks.Visual and longer CoT can accelerate convergence but do not substantially raise the final performance ceiling.Source: §6 · p. 10
- Concise, well-grounded supervision appears more effective for learning reusable visual reasoning patterns than verbose, heavily supervised traces.The paper identifies this as a “short is long” effect and plans to extend the analysis to richer task families beyond mazes and VLMs.Source: §6 · p. 10
Data Synthesis
The data synthesis section describes three CoT formats generated from maze paths: language directions, grounding coordinates, and visual traces. Each format uses a rule-based path transformation and a Gemini-2.5-Pro prompt to synthesize reasoning trajectories.
Language CoT Synthesis
- Language CoT synthesis converts a maze path into stepwise directions, then prompts Gemini-2.5-Pro to produce reasoning trajectories.The detailed prompt is shown in Figure 6.Source: Data Synthesis · p. 15, Figure 6 · p. 16
Grounding CoT Synthesis
- Grounding CoT synthesis maps each grid visited along the path to its absolute image coordinates before prompting Gemini-2.5-Pro.The detailed prompt is shown in Figure 7.Source: Data Synthesis · p. 15, Figure 7 · p. 17
Visual CoT Synthesis
- Visual CoT synthesis draws a line through the path’s image coordinates, then prompts Gemini-2.5-Pro to produce reasoning trajectories.The prompt is shown in Figure 8.Source: Data Synthesis · p. 15, Figure 8 · p. 18