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

TL;DR: In vision-centric reasoning (e.g., mazes, puzzles), "Short is Long". Our research reveals that concise, coordinate-based reasoning generalizes significantly better than verbose language or visual manipulation CoT chains.

---

📖 Overview

Recent advancements in Vision-Language Models (VLMs) often rely on lengthy Chain-of-Thought (CoT) or "Thinking with Images" (Visual CoT) to improve reasoning. However, does more verbose or visual reasoning actually lead to better generalization?

In this work, we systematically evaluate different CoT designs using a controlled maze-solving benchmark (and validate on real-world tasks). We utilize a SFT-then-RL (Supervised Fine-Tuning followed by Reinforcement Learning) pipeline based on Qwen2.5-VL-7B.

Key Findings

1. Visual CoT Accelerates, but Doesn't Elevate: Methods that "draw" on images (Visual CoT) speed up convergence during RL but do not raise the final performance ceiling compared to simpler methods.
2. Concise Grounding is Better: A minimal trajectory of coordinates (Implicit CoT / G-CoT-least) outperforms verbose step-by-step language reasoning.
3. The "Short is Long" Effect: The most concise formats (containing only essential grounding/path info) achieve the best generalization across different scales (e.g., training on $6\times6$ mazes and solving $7\times7$ mazes).

---

🧠 CoT Paradigms Compared

We isolate and compare three distinct ways of externalizing reasoning:

Method	Description
Language CoT	Pure text reasoning (e.g., "Move North, then West...")
Grounding CoT	Text linked with coordinates (e.g., "Move to [x,y]...")
Visual CoT	Interleaved image manipulation (drawing lines/marks)
G-CoT-least	Implicit sequence of coordinates/path only.

---

📊 Experimental Highlights

1. Visual CoT Improves Efficiency, Not Efficacy

Visual CoT does not offer a higher performance upper bound compared to Language or Grounding CoT, but only accelerates training.

2. Short CoT Surpasses Longer Ones

The VLM can conduct implicit reasoning after its grounding ability is aligned with the visual environment, achieving full accuracy and faster convergence.

3. Least CoT Achieves Better Generalization

Properly aligned grounding ability allows the model to generalize its spatial reasoning effectively to new visual environments.

4. Beyond Mazes

We validated our "Short is Long" hypothesis on broader vision-centric tasks:

• Visual Games: FrozenLake, Jigsaw Puzzles.
• Real-world VQA: $V^*$ Bench, HR-Bench.
• Result: G-CoT-least consistently achieved state-of-the-art results compared to verbose CoT methods.

Evaluation results on other vision reasoning tasks, where the best-performed results are marked as bold.

Model	$V^*$ Bench			HR-Bench 4K			Frozenlake	Jigsaw
	Attr	Spatial	Overall	FSP	FCP	Overall
Qwen2.5-VL-7B	67.83	78.95	72.25	88.00	57.00	72.50	20.00	0.00
+ V-CoT RL	86.09	78.95	83.25	87.00	57.00	72.00	-	-
+ G-CoT-least RL	87.83	82.89	85.86	90.75	57.50	74.12	90.33	75.60

---

citation

If you find this work helpful for your research, please consider citing: