Research map / Vision-language and multimodal / Multimodal LLMs
Chain-of-Visual-Thought: Teaching VLMs to See and Think Better with Continuous Visual Tokens
Yiming Qin, Bomin Wei, Jiaxin Ge, Konstantinos Kallidromitis, Stephanie Fu, Trevor Darrell, XuDong Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CoVT with 3 visual tokens | Qwen2.5-VL-7B · CV-Bench | accuracy: 74.5 -> 80.0 | Improved |
| add CoVT with 3 visual tokens | Qwen2.5-VL-7B · CV-Bench Depth | accuracy: 72.8 -> 86.8 | Improved |
| add CoVT with 3 visual tokens | Qwen2.5-VL-7B · MME-RealWorld | accuracy: 68.6 -> 72.9 | Improved |
| add CoVT with 3 visual tokens | Qwen2.5-VL-7B · HRBench-8K | accuracy: 64.9 -> 69.4 | Improved |
| add edge token to make four visual tokens | CoVT · CV-Bench Depth | accuracy: 86.8 -> 89.2 | Improved |
| add edge token to make four visual tokens | CoVT · MMVP | accuracy: 58.7 -> 56.7 | Hurt |
| add segmentation tokens (8) instead of 0 token fine-tuning (ablation) | CoVT · CV-Bench | accuracy: 76.6 -> 80.0 | Improved |
| use 16 empty tokens without visual alignment (ablation) | CoVT · CV-Bench | accuracy: 76.6 -> 75.7 | Hurt |
| increase segmentation tokens from 8 to 32 (ablation) | CoVT · CV-Bench | accuracy: 80.0 -> 73.9 | Hurt |
| use only 1 segmentation token instead of 8 (ablation) | CoVT · CV-Bench | accuracy: 80.0 -> 78.9 | Hurt |
| align tokens through task decoders instead of direct feature MSE (ablation) | CoVT Seg · BLINK | accuracy: 55.2 -> 57.4 | Improved |
| align tokens through task decoders instead of direct feature MSE (ablation) | CoVT Depth · CV-Bench | accuracy: 77.0 -> 78.7 | Improved |
Closest papers by meaning
- VLM-R³: Region Recognition, Reasoning, and Refinement for Enhanced Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.92)
- VLMs have Tunnel Vision: Evaluating Nonlocal Visual Reasoning in Leading VLMs (NeurIPS 2025 · similarity 0.92)
- CoFFT: Chain of Foresight-Focus Thought for Visual Language Models (NeurIPS 2025 · similarity 0.92)
- Visual Thoughts: A Unified Perspective of Understanding Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.92)
- VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.92)
- Perception Tokens Enhance Visual Reasoning in Multimodal Language Models (CVPR 2025 · similarity 0.91)
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.
