Research map / Vision-language and multimodal / Multimodal LLMs
Revisiting the Necessity of Lengthy Chain-of-Thought in Vision-centric Reasoning Generalization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Compares language, grounding, and visual chain-of-thought formats in an SFT-then-RL pipeline on Qwen2.5-VL-7B using maze navigation, finding visual CoT speeds convergence but not the ceiling and the least-grounded CoT generalizes best.
Problem
It is unclear why specific chain-of-thought designs (language, grounding, visual) help VLMs acquire generalizable vision-centric reasoning, and which CoT forms truly support generalization.
Approach
Uses a contamination-free maze testbed to control CoT format while holding model and training fixed, showing that short, grounded CoT (sparse coordinate paths) generalizes better across maze sizes than verbose language or visual CoT, and that visual CoT mainly accelerates training.
Key results
- Visual CoT RL converges in roughly half the steps of language CoT but does not exceed grounding CoT's final training accuracy ceiling.
- G-CoT-least reaches 85.86 overall on V* Bench vs 72.25 for Qwen2.5-VL-7B and 83.25 for V-CoT RL.
- On Jigsaw, G-CoT-least raises accuracy from 0.00 to 75.60, and on FrozenLake from 20.00 to 90.33.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| visual CoT (V-CoT) SFT then RL vs language/grounding CoT (ablation) | Qwen2.5-VL-7B maze navigation · Mazes 4x4-6x6 train, 7x7 test | RL convergence speed and final accuracy | Mixed |
| shorter least grounding CoT (path only) instead of explicit grounding CoT (ablation) | Qwen2.5-VL-7B maze navigation · Mazes 4x4-6x6 train | training convergence and starting accuracy | Helped |
| least grounding CoT (G-CoT-least) instead of visual CoT (ablation) | Qwen2.5-VL-7B maze navigation · Cross-scale: SFT 4x4-6x6, RL 7x7-9x9, test 10x10 | test accuracy on unseen mazes | Helped |
| add least grounding CoT RL (vs base) on V* Bench | Qwen2.5-VL-7B · V* Bench overall | accuracy: +13.61 over base (72.25 to 85.86) | Helped |
| least grounding CoT RL vs visual CoT RL on V* Bench | Qwen2.5-VL-7B · V* Bench overall | accuracy: 85.86 vs 83.25 | Helped |
| least grounding CoT RL vs visual CoT RL on HR-Bench 4K | Qwen2.5-VL-7B · HR-Bench 4K overall | accuracy: 74.12 vs 72.00 | Helped |
| least grounding CoT RL on Jigsaw | Qwen2.5-VL-7B · Jigsaw 3x3 | accuracy: 0.00 to 75.60 | Helped |
| least grounding CoT RL on FrozenLake | Qwen2.5-VL-7B · FrozenLake 4x4 | success rate: 20.00 to 90.33 | Helped |
| visual CoT RL vs grounding on HR-Bench 4K | Qwen2.5-VL-7B · HR-Bench 4K overall | accuracy: -0.50 vs base (72.50 to 72.00) | Hurt |
| visual CoT RL on V* Bench | Qwen2.5-VL-7B · V* Bench overall | accuracy: +11.00 over base (72.25 to 83.25) | Helped |
Surprising findings
- Visual CoT, despite richer supervision, does not raise the final performance ceiling over shorter CoT.
- Least grounded CoT (sparse coordinate path with no explicit coordinates in text) converges faster than explicit grounding CoT and still reaches 100% training accuracy.
Limitations
- Main experiments limited to a synthetic maze testbed and a few additional tasks.
- Relies on a single base model (Qwen2.5-VL-7B).
- Visual CoT uses predefined operations (line drawing, cropping), which may not represent all visual CoT designs.
Experimental setup
Datasets: synthetic maze dataset (4x4 to 9x9) Jigsaw dataset from prior work V* training set
Benchmarks: maze navigation (7x7, 10x10 test) V* Bench HR-Bench 4K FrozenLake Jigsaw
Metrics: success rate / accuracy pass@8 training accuracy
Baselines: Language CoT (L-CoT) Grounding CoT (G-CoT) Visual CoT (V-CoT) Qwen2.5-VL-7B zero RL
Builds on: GRPO Chain-of-Thought DeepEyes verl LLaMAFactory
Closest papers by meaning
- Zebra-CoT: A Dataset for Interleaved Vision-Language Reasoning (ICLR 2026 · similarity 0.91)
- From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models (ICML 2026 · similarity 0.91)
- HopChain: Multi-Hop Data Synthesis for Generalizable Vision-Language Reasoning (NeurIPS 2026 · similarity 0.91)
- Unveiling Chain of Step Reasoning for Vision-Language Models with Fine-grained Rewards (NeurIPS 2025 · similarity 0.90)
- Uni-CoT: Towards Unified Chain-of-Thought Reasoning Across Text and Vision (ICLR 2026 · similarity 0.90)
- VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.