Research map / Vision-language and multimodal / Multimodal LLMs
SRUM: Fine-Grained Self-Rewarding for Unified Multimodal Models
Weiyang Jin, Yuwei Niu, Jiaqi Liao, Chengqi Duan, Aoxue Li, Shenghua Gao, Xihui Liu
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 SRUM self-rewarding post-training | Bagel (CoT) · T2I-CompBench (Qwen-evaluated) | overall score: 84.46 -> 88.37 | Improved |
| add SRUM self-rewarding post-training | Bagel · T2I-CompBench (Qwen-evaluated) | overall score: 82.18 -> 86.55 | Improved |
| add SRUM self-rewarding post-training | BLIP3o · T2I-CompBench (Qwen-evaluated) | overall score: 84.66 -> 85.75 | Improved |
| add local reward (ablation) | SRUM (Bagel) · T2I-CompBench | overall score: 0.76 drop in CoT mode, 1.04 drop in standard mode when omitted | Improved |
| add global reward (ablation) | SRUM (Bagel) · T2I-CompBench | overall score | Improved |
| add reference constraint (ablation) | SRUM (Bagel) · T2I-CompBench | overall score | Improved |
| use dense continuous reward instead of sparse 0-1 reward (ablation) | SRUM (Bagel) · T2I-CompBench | overall score | Improved |
| set reference constraint weight lambda_c to 0.5 (ablation) | SRUM (Bagel) · T2I-CompBench | overall score | Improved |
| use internal understanding module instead of Qwen 32B external reward (ablation) | SRUM (Bagel) · T2I-CompBench (external reward comparison) | overall score: 88.01 -> 88.37 | Improved |
| add global reward to Qwen 7B external reward (ablation) | Bagel · T2I-CompBench (external reward comparison) | overall score: 85.78 -> 87.52 | Improved |
| add global reward to Qwen 32B external reward (ablation) | Bagel · T2I-CompBench (external reward comparison) | overall score: 85.53 -> 88.01 | Improved |
| add SRUM reward-weighted training over plain SFT on self-generated data | Bagel · T2I-ReasonBench | overall accuracy: 45.37 -> 46.75 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- OmniGen2 — OmniGen2: Towards Instruction-Aligned Multimodal Generation (CVPR 2026) · better on 1 of 1 shared result
- Show-O2 — Show-o2: Improved Native Unified Multimodal Models (NeurIPS 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Learning to Generate via Understanding: Understanding-Driven Intrinsic Rewarding for Unified Multimodal Models (CVPR 2026 · similarity 0.92)
- Calibrated Self-Rewarding Vision Language Models (NeurIPS 2024 · similarity 0.90)
- ARM-Thinker: Reinforcing Multimodal Generative Reward Models with Agentic Tool Use and Visual Reasoning (CVPR 2026 · similarity 0.90)
- SSL4RL: Revisiting Self-supervised Learning as Intrinsic Reward for Visual-Language Reasoning (ICML 2026 · similarity 0.89)
- VL-RewardBench: A Challenging Benchmark for Vision-Language Generative Reward Models (CVPR 2025 · similarity 0.89)
- Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and Generation (CVPR 2025 · similarity 0.89)
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.
