Research map / Vision-language and multimodal / Multimodal LLMs
Structured Policy Optimization: Enhance Large Vision-Language Model via Self-referenced Dialogue
Guohao Sun, Can Qin, Yihao Feng, Zeyuan Chen, Ran Xu, Sohail Dianat, Majid Rabbani, Raghuveer Rao, Zhiqiang Tao
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 question reward r(x,q) to SPO (ablation) | SFT-7B-SPO · LLaVA-Bench-Wilder | accuracy: 69.7 -> 70.2 | Improved |
| add answer reward r([x,q],y) to SPO (ablation) | SFT-7B-SPO · MME | score: 1571 -> 1585 | Improved |
| add both question and answer rewards (ablation) | SFT-7B · MuirBench | accuracy: 45.5 -> 46.7 | Improved |
| add question reward (ablation) | SFT-7B-SPO · MLVU | accuracy: 65.0 -> 65.3 | No clear effect |
| use SPO instead of DPO | SFT-0.5B · MMVet | accuracy: 25.8 -> 30.0 | Improved |
| use SPO instead of DPO | SFT-7B · MMStar | accuracy: 61.9 -> 62.1 | No clear effect |
| use DPO on SFT-0.5B | SFT-0.5B · LLaVA-Bench-Wilder | accuracy: 59.4 -> 58.5 | Hurt |
| use SPO instead of DPO | SFT-7B · EgoSchema | accuracy: 59.5 -> 60.5 | Improved |
| use SPO instead of DPO | SFT-7B · LLaVA-W 60 questions | inference time (s): 547 -> 254 | Improved |
| use SPO instead of DPO | SFT-0.5B, High-Q MMVet split · MMVet | accuracy drop w/o image: 77.8% -> 72.4% | Improved |
| use SPO instead of DPO | SFT-0.5B, Low-Q MMVet split · MMVet | accuracy drop w/o image: 64.7% -> 47.8% | Improved |
| use SPO instead of DPO | LLaVA-OV training · training time | hours (4xA100): 22h -> 24h | No clear effect |
Closest papers by meaning
- DRESS: Instructing Large Vision-Language Models to Align and Interact with Humans via Natural Language Feedback (CVPR 2024 · similarity 0.89)
- Self-alignment of Large Video Language Models with Refined Regularized Preference Optimization (NeurIPS 2025 · similarity 0.89)
- Sequential Preference Optimization: Multi-Dimensional Preference Alignment with Implicit Reward Modeling (AAAI 2025 · similarity 0.88)
- 3D-Properties: Identifying Challenges in DPO and Charting a Path Forward (ICLR 2025 · similarity 0.88)
- Multi-Reference Preference Optimization for Large Language Models (AAAI 2025 · similarity 0.88)
- MMDU: A Multi-Turn Multi-Image Dialog Understanding Benchmark and Instruction-Tuning Dataset for LVLMs (NeurIPS 2024 · similarity 0.88)
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.
