Research map / Vision-language and multimodal / Multimodal LLMs

Uni-Synergy: Bridging Understanding and Generation for Personalized Reasoning

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning

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 Sync-R1, an RL framework (Sync-GRPO with Dynamic Group Scaling) that jointly optimizes personalized understanding and image generation in a unified model, and a harder benchmark UnifyBench++.

Problem

Unified multimodal models struggle to link personalized understanding and generation; prior work relies on implicit token-level alignment via SFT, which under-exploits synergy between comprehension and creation.

Approach

Orchestrates an explicit two-phase reasoning trajectory (understanding reasoning then context-guided generation) optimized end-to-end with a GRPO variant (Sync-GRPO) using an ensemble reward, plus Dynamic Group Scaling that prunes low-potential trajectories early using an intermediate surrogate reward and an adaptive threshold.

Key results

• Sync-R1 (1.3B) reports 0.859 recognition weight and 0.250 Rea. BLEU, beating UniCTokens (0.792, 0.238) and GPT-4o+IP on several understanding metrics.
• Relative improvements over UniCTokens of 14.1% in Dense Gen., 10.0% in Rea. Gen., and 8.8% in Dense Rea. Gen.
• DGS reported to reach 98% of baseline peak performance about 1.9x faster in wall-clock time.

Surprising findings

• Sync-GRPO applied on a simple Joint Training initialization gives gains comparable to or above the complex UniCTokens pipeline, suggesting cold-start procedures are not essential.

Limitations

• Relies on Show-o 1.3B backbone and a multi-stage UniCTokens initialization in main experiments.
• Theoretical variance-reduction result rests on strong bivariate normal and i.i.d. assumptions and is loosely justified.
• Key metrics (Dense Rea., Dense Gen.) depend on GPT-4o judging, which is not validated with human evaluation.

Experimental setup

Datasets: UnifyBench UnifyBench++

Benchmarks: UnifyBench UnifyBench++

Metrics: BLEU GPT-4o score CLIP-T CLIP-I recognition weight

Baselines: UniCTokens Yo'Chameleon Yo'LLaVA MC-LLaVA RAP-MLLM Chameleon Show-o DreamBooth Textual Inversion GPT-4o

Builds on: GRPO Dr.GRPO UniCTokens Show-o MaskGIT

Closest papers by meaning

• Diving into Self-Evolving Training for Multimodal Reasoning (ICML 2025 · similarity 0.87)
• R-C2: Cycle-Consistent Reinforcement Learning Improves Multimodal Reasoning (CVPR 2026 · similarity 0.87)
• Co-Reinforcement Learning for Unified Multimodal Understanding and Generation (NeurIPS 2025 · similarity 0.87)
• Learning to Generate via Understanding: Understanding-Driven Intrinsic Rewarding for Unified Multimodal Models (CVPR 2026 · similarity 0.87)
• Boosting LLM Reasoning via Human-Inspired Reward Shaping (NeurIPS 2026 · similarity 0.87)
• Revisual-R1: Advancing Multimodal Reasoning From Optimized Cold Start to Staged Reinforcement Learning (ICLR 2026 · similarity 0.87)