Research map / Vision-language and multimodal / Multimodal LLMs

Learning to Generate via Understanding: Understanding-Driven Intrinsic Rewarding for Unified Multimodal Models

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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 GvU, a self-supervised GRPO-based RL method where a unified multimodal model (X-Omni) uses its own understanding branch's token-level text-image likelihood as an intrinsic reward to improve text-to-image generation.

Problem

Unified multimodal models understand images better than they generate them; the understanding-generation gap limits T2I quality, and existing rewards are coarse or prone to hacking.

Approach

Use the UMM's own understanding branch to compute the geometric-mean token probability of the ground-truth prompt given the generated image, and use this token-level intrinsic likelihood as the GRPO reward, with no external reward model or image data.

Key results

• GenEval overall improves from 0.68 to 0.81 without LLM rewriting (0.84 with rewriter) after GvU RL.
• GenEval++ overall rises from 0.282 to 0.404 (43.3% relative gain), the top score in its comparison table.
• DPG-Bench overall reaches 85.68, the best among compared models; on MMT-Bench overall the base improves from 49.76 to 49.92, with larger gains on fine-grained subtasks such as VI (+5.06).

Surprising findings

• Training from a weaker base model yields a much larger relative gain on GenEval (+138.1% vs +19.1% for the regular base), suggesting a larger understanding-generation gap helps.
• Gains on understanding benchmarks are modest, and the method is trained without any image-understanding supervision yet still improves some fine-grained MMT-Bench subtasks.

Limitations

• Understanding gains are relatively modest.
• Results reported primarily on a single base model (X-Omni) with a self-built RL framework.
• Reward is a likelihood proxy; sensitivity analysis is only qualitative on prompt-term removal.

Experimental setup

Datasets: self-generated text prompt dataset (50,000 samples)

Benchmarks: GenEval DPG-Bench GenEval++ POPE GQA MMBench SEED DocVQA OCRBench MMT-Bench

Metrics: overall score accuracy

Baselines: SD3-medium FLUX.1-dev BAGEL Janus-Pro OmniGen2 XOmni GPT-4o BLIP3-o 8B UniWorld-V1

Builds on: GRPO X-Omni TRL LoRA

Closest papers by meaning

• Generation Enhances Understanding in Unified Multimodal Models via Multi-Representation Generation (ICML 2026 · similarity 0.90)
• Towards Unified Multimodal Interleaved Generation via Group Relative Policy Optimization (NeurIPS 2025 · similarity 0.90)
• Janus-Pro-R1: Advancing Collaborative Visual Comprehension and Generation via Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
• DuoGen: Towards Autonomous Interleaved Multimodal Generation (CVPR 2026 · similarity 0.90)
• Vision as a Dialect: Unifying Visual Understanding and Generation via Text-Aligned Representations (NeurIPS 2025 · similarity 0.89)
• Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and Generation (CVPR 2025 · similarity 0.89)