Short Is Long: What Kind of Chain-of-Thought Actually Teaches a VLM to Reason?

2026-07-13 · Source: LLM on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A CVPR 2026 paper by Renmin's GSAI and ByteDance Seed challenges the assumption that longer, more detailed chain-of-thought (CoT) data is superior for teaching Visual Language Models (VLMs) to reason. Using maze solving as a testbed, where models like Qwen2.5-VL-7B achieve under 10% success on 4x4 grids, the researchers compared Language CoT, Grounding CoT (coordinate trajectories), and Visual CoT (image manipulation) within an SFT-then-RL pipeline. Their findings indicate that while Visual CoT converges faster than Language CoT, it does not achieve higher accuracy. Crucially, a "short CoT" approach, which uses sparse coordinates or final trajectories without verbose narration, proved most efficient, starting higher, converging faster, and outperforming Visual CoT. This minimal grounding CoT also demonstrated superior generalization to larger, unseen maze sizes like 7x7 and 10x10, supporting the "short is long" principle.

Key takeaway

For Machine Learning Engineers developing Visual Language Models for complex reasoning, you should re-evaluate the common assumption that more verbose chain-of-thought (CoT) data is always better. Instead, focus on generating or curating "short CoT" data, utilizing sparse coordinates or final trajectories. This approach, demonstrated to be more efficient and generalize better to unseen problem sizes, can significantly improve your model's reasoning capabilities and training efficiency. Prioritize quality and conciseness in your CoT data over sheer volume or granularity.

Key insights

Short, sparse chain-of-thought data significantly improves VLM reasoning efficiency and generalization.

Principles

• Less verbose CoT improves VLM reasoning.
• Efficiency doesn't always mean higher accuracy.
• Minimal grounding enhances generalization.

Method

Compared Language, Grounding, and Visual CoT via SFT-then-RL pipeline on maze solving tasks.

In practice

• Prioritize sparse CoT over verbose narration.
• Test CoT methods on generalization to unseen scales.
• Consider maze solving as a VLM reasoning benchmark.

Topics

• Visual Language Models
• Chain-of-Thought
• VLM Reasoning
• Supervised Fine-Tuning
• Reinforcement Learning
• Generalization

Best for: Research Scientist, Computer Vision Engineer, AI Scientist, Machine Learning Engineer

Related on AIssential

• Thinking in Pictures: A Diagnostic Study of Visual vs. Textual Chain-of-Thought Reasoning in Vision-Language Models — Visual Chain-of-Thought significantly outperforms textual CoT for inherently visual tasks in Vision-Language Models.
• Teaching Reasoning Models When to Stop: Data, Rewards, and Self-Aware Decoding — Efficient CoT reasoning requires balancing length reduction with accuracy, often by leveraging model confidence.
• When Thinking Hurts: Epistemic Signals in the Reasoning Chains of Visual Language Models — Reasoning chain entropy is a more reliable uncertainty signal than answer entropy in thinking-mode VLMs.
• CoT-Space: A Theoretical Framework for Internal Slow-Thinking via Reinforcement Learning — CoT-Space reframes LLM reasoning as continuous optimization in a semantic space, explaining optimal CoT length.
• Answer-Conditioned Chain-of-Thought Distillation for Few-Shot Industrial Vision with Small VLMs — Answer-conditioned CoT distillation significantly improves small VLM performance for few-shot industrial vision by ensuring correct reasoning.
• Do Thoughts Depth Affect Multilingual Reasoning? — Chain-of-Thought effectiveness in multilingual LLMs depends on reasoning depth and language resource availability.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by LLM on Medium.