Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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 how different CoT designs—linguistic, grounding-based, and visual—affect the acquisition and generalization of visual reasoning in vision-language models. Using a controlled maze benchmark and a robust SFT-then-RL pipeline on Qwen2.5-VL-7B, it reveals that visual and longer CoTs speed up training but do not raise final performance, while concise grounding-focused CoTs (especially G-CoT-least) yield the strongest cross-size generalization. The results extend to visual games and real-world VQA, where minimal grounding proves most effective, suggesting a universal 'short is long' principle for generalizable visual reasoning. The findings offer practical guidelines for building generalizable SFT datasets and highlight implicit spatial grounding as a core driver of transfer in vision-centric tasks.

Abstract

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.

Revisiting the Necessity of Lengthy Chain-of-Thought in Vision-centric Reasoning Generalization

TL;DR

The paper investigates how different CoT designs—linguistic, grounding-based, and visual—affect the acquisition and generalization of visual reasoning in vision-language models. Using a controlled maze benchmark and a robust SFT-then-RL pipeline on Qwen2.5-VL-7B, it reveals that visual and longer CoTs speed up training but do not raise final performance, while concise grounding-focused CoTs (especially G-CoT-least) yield the strongest cross-size generalization. The results extend to visual games and real-world VQA, where minimal grounding proves most effective, suggesting a universal 'short is long' principle for generalizable visual reasoning. The findings offer practical guidelines for building generalizable SFT datasets and highlight implicit spatial grounding as a core driver of transfer in vision-centric tasks.

Abstract

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.

Paper Structure

This paper contains 31 sections, 3 equations, 8 figures, 1 table.

Table of Contents

  1. Introduction
  2. Preliminary
  3. Language CoT
  4. Grounding CoT
  5. Visual CoT
  6. Experimental Setup
  7. Maze Data Construction
  8. Language CoT Synthesis.
  9. Grounding CoT Synthesis.
  10. Visual CoT Synthesis.
  11. Training Strategy
  12. Supervised Fine-Tuning.
  13. Reinforcement Learning.
  14. Implementation Details
  15. Experimental Analysis
  16. ...and 16 more sections

Figures (8)

  • Figure 1: Illustration of four types of CoT reasoning strategies. We show examples of solving a 6$\times$6 maze navigation task.
  • Figure 2: Training dynamics of zero RL, Grounding CoT (G-CoT), Languge CoT (L-CoT), and Visual CoT (V-CoT). The model is cold-started on mazes sized $4\times4$ to $6\times6$.
  • Figure 3: Training Dynamics of least Grounding CoT (G-CoT-least), Grounding CoT (G-CoT), and Visual CoT (V-CoT). The model is cold-started on mazes sized $4\times4$ to $6\times6$.
  • Figure 4: Training Dynamics of least Grounding CoT (G-CoT-least) and Visual CoT (V-CoT). The model is cold-started on mazes sized $6\times6$ to validate the single-scale generalization.
  • Figure 5: Training Dynamics of least Grounding CoT (G-CoT-least) and Visual CoT (V-CoT). The model is cold-started on mazes sized $4\times4$ to $6\times6$ to validate the cross-scale generalization.
  • ...and 3 more figures