Research map / Vision-language and multimodal / Multimodal LLMs
Self-Rewarding Vision-Language Model via Reasoning Decomposition and Multi-Reward Policy Optimization
Zongxia Li, Wenhao Yu, Chengsong Huang, Zhenwen Liang, Rui Liu, Fuxiao Liu, Jingxi Chen, Dian Yu, Jordan Lee Boyd-Graber, Haitao Mi, Dong Yu
Read the paperCheck your idea against 137,217 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.
Proposes a self-rewarding post-training framework for vision-language models that decomposes reasoning into visual perception and language reasoning and optimizes multiple rewards to curb hallucinations and language shortcuts.
Problem
VLMs suffer from visual hallucinations and language shortcuts because post-training relies on verifiable answer matching and supervises only final outputs, leaving intermediate visual reasoning unguided.
Approach
Decomposes VLM reasoning into explicit visual perception and language reasoning stages, and uses self-generated multi-reward signals (no external reward model implied in abstract) within policy optimization to supervise intermediate visual reasoning rather than only final answers.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add visual self-reward and multi-reward policy optimization (ablation) | Vision-SR1 (Qwen2.5-VL-7B) · 7-benchmark average | accuracy: 50.3 -> 52.2 | Improved |
| add visual self-reward and multi-reward policy optimization (ablation) | Vision-SR1 (Qwen2.5-VL-3B) · 7-benchmark average | accuracy: 47.6 -> 48.8 | Improved |
| add visual self-reward (ablation) | Vision-SR1 (7B) · MMMU-Pro | accuracy: 42.8 -> 40.7 | Hurt |
| add visual self-reward (ablation) | Vision-SR1 (7B) · VisNumBench | accuracy: 35.7 -> 43.5 | Improved |
| use Vision-SR1 instead of answer-reward-only Vision-R1 on same 47K data | Qwen2.5-VL-7B · 7-benchmark average | accuracy: 50.7 -> 52.2 | Improved |
| use Vision-SR1 instead of Vision-R1 | Mimo-VL-7B · 7-benchmark average | accuracy: 46.0 -> 49.5 | Improved |
| use Vision-SR1 instead of Vision-R1 | Qwen2.5-VL-3B · MATH-Vision | accuracy: 29.9 -> 29.3 | No clear effect |
| use Vision-SR1 instead of Vision-R1 | Qwen2.5-VL-7B · OmniSpatial | accuracy: 31.1 -> 44.2 | Improved |
| use Vision-SR1 instead of Vision-R1 | Qwen2.5-VL-7B · ViLP (language shortcut) | accuracy: 51.3 -> 52.6 | Improved |
| add self-reward to reduce language shortcut rate (ablation) | Vision-SR1 (3B) · avg LSR over 7 benchmarks | LSR: 10.4 -> 9.4 | Improved |
| add self-reward to reduce language shortcut rate (ablation) | Vision-SR1 (7B) · avg LSR over 7 benchmarks | LSR: 10.1 -> 9.8 | No clear effect |
| use multimodal RL (Vision-SR1) on text-only math | Qwen2.5-VL-7B · MATH-500 | accuracy: 73.4 -> 70.8 | Hurt |
Closest papers by meaning
- SSL4RL: Revisiting Self-supervised Learning as Intrinsic Reward for Visual-Language Reasoning (ICML 2026 · similarity 0.94)
- PDCR: Perception-Decomposed Confidence Reward for Vision-Language Reasoning (CVPR 2026 · similarity 0.93)
- Calibrated Self-Rewarding Vision Language Models (NeurIPS 2024 · similarity 0.92)
- VL-RewardBench: A Challenging Benchmark for Vision-Language Generative Reward Models (CVPR 2025 · similarity 0.92)
- From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models (ICML 2026 · similarity 0.92)
- Through the Lens of Contrast: Self-Improving Visual Reasoning in VLMs (ICLR 2026 · 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.
