AI Paper Atlas

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

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis

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 changedWhereMeasuredEffect
add visual self-reward and multi-reward policy optimization (ablation)Vision-SR1 (Qwen2.5-VL-7B) · 7-benchmark averageaccuracy: 50.3 -> 52.2Improved
add visual self-reward and multi-reward policy optimization (ablation)Vision-SR1 (Qwen2.5-VL-3B) · 7-benchmark averageaccuracy: 47.6 -> 48.8Improved
add visual self-reward (ablation)Vision-SR1 (7B) · MMMU-Proaccuracy: 42.8 -> 40.7Hurt
add visual self-reward (ablation)Vision-SR1 (7B) · VisNumBenchaccuracy: 35.7 -> 43.5Improved
use Vision-SR1 instead of answer-reward-only Vision-R1 on same 47K dataQwen2.5-VL-7B · 7-benchmark averageaccuracy: 50.7 -> 52.2Improved
use Vision-SR1 instead of Vision-R1Mimo-VL-7B · 7-benchmark averageaccuracy: 46.0 -> 49.5Improved
use Vision-SR1 instead of Vision-R1Qwen2.5-VL-3B · MATH-Visionaccuracy: 29.9 -> 29.3No clear effect
use Vision-SR1 instead of Vision-R1Qwen2.5-VL-7B · OmniSpatialaccuracy: 31.1 -> 44.2Improved
use Vision-SR1 instead of Vision-R1Qwen2.5-VL-7B · ViLP (language shortcut)accuracy: 51.3 -> 52.6Improved
add self-reward to reduce language shortcut rate (ablation)Vision-SR1 (3B) · avg LSR over 7 benchmarksLSR: 10.4 -> 9.4Improved
add self-reward to reduce language shortcut rate (ablation)Vision-SR1 (7B) · avg LSR over 7 benchmarksLSR: 10.1 -> 9.8No clear effect
use multimodal RL (Vision-SR1) on text-only mathQwen2.5-VL-7B · MATH-500accuracy: 73.4 -> 70.8Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy