Research map / Vision-language and multimodal / Multimodal LLMs
Unsupervised Visual Chain-of-Thought Reasoning via Preference Optimization
Kesen Zhao, Beier Zhu, Qianru Sun, Hanwang Zhang
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 image-level CoT inference with generated bounding boxes (ablation) | UV-CoT · Six VQA benchmarks average | GPT-4o score (0-1): 0.417 -> 0.494 | Improved |
| use ground-truth bounding boxes instead of predicted ones (ablation) | UV-CoT · Six VQA benchmarks average | GPT-4o score (0-1): 0.494 -> 0.618 | Improved |
| replace naive DPO with Score-DPO (sDPO) (ablation) | UV-CoT · Six VQA benchmarks average | GPT-4o score (0-1): 0.475 -> 0.494 | Improved |
| add iterative learning (ablation) | UV-CoT · Six VQA benchmarks average | GPT-4o score (0-1): 0.459 -> 0.494 | Improved |
| add next-response term gamma in evaluation (ablation) | UV-CoT · Six VQA benchmarks average | GPT-4o score (0-1): 0.406 -> 0.494 | Improved |
| use OmniLMM-12B evaluator instead of self-evaluation (ablation) | UV-CoT · Six VQA benchmarks average | GPT-4o score (0-1): 0.441 -> 0.473 | Improved |
| add UV-CoT bounding-box CoT at inference | OmniLMM-12B · Six VQA benchmarks average | GPT-4o score (0-1): 0.443 -> 0.516 | Improved |
| add UV-CoT bounding-box CoT at inference | LLaVA-1.5-7B · Six VQA benchmarks average | GPT-4o score (0-1): 0.393 -> 0.440 | Improved |
| add 10% labeled SFT before preference optimization | UV-CoT · Six VQA benchmarks average | GPT-4o score (0-1): 0.473 -> 0.494 | Improved |
| use unsupervised UV-CoT instead of supervised Visual-CoT | LLaVA-1.5-7B · Zero-shot DUDE/SROIE/Visual7w average | GPT-4o score (0-1): 0.261 -> 0.286 | Improved |
| train on unlabeled preference data from zero-shot datasets (UV-CoT*) | UV-CoT · Zero-shot DUDE/SROIE/Visual7w average | GPT-4o score (0-1): 0.286 -> 0.312 | Improved |
| use UV-CoT instead of Visual-CoT-7B | LLaVA-1.5-7B · V* Bench zero-shot average | GPT-4o score (0-1): 0.347 -> 0.402 | Improved |
Closest papers by meaning
- TVI-CoT: Text-Visual Interleaved Chain-of-Thought Reasoning for Multimodal Understanding (ICML 2026 · similarity 0.93)
- Uni-CoT: Towards Unified Chain-of-Thought Reasoning Across Text and Vision (ICLR 2026 · similarity 0.93)
- Visual CoT: Advancing Multi-Modal Language Models with a Comprehensive Dataset and Benchmark for Chain-of-Thought Reasoning (NeurIPS 2024 · similarity 0.93)
- Zebra-CoT: A Dataset for Interleaved Vision-Language Reasoning (ICLR 2026 · similarity 0.92)
- Revisiting the Necessity of Lengthy Chain-of-Thought in Vision-centric Reasoning Generalization (CVPR 2026 · similarity 0.91)
- Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 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.
