The Year in AI Papers
The year/Independent research

Paper 2603.03241

UniG2U-Bench: Do Unified Models Advance Multimodal Understanding?

Zimo Wen, Boxiu Li, Wanbo Zhang, Junxiang Lei, Xiaoyu Chen, Yijia Fan, Qi Zhang, Yujiang Wang, Lili Qiu, Bo Li, Ziwei Liu, Caihua Shan, et al.

Published
Mar 2026
Research lab
Independent
Citations
3
GitHub
Not linked

01 In brief

Summary

UniG2U-Bench is a new benchmark for evaluating whether unified multimodal models (UMMs) benefit from generation when performing understanding tasks.

It includes 3,000 samples across 7 categories and 30 subtasks, and evaluates over 30 models, including base VLMs, unified models, and agentic models.

The study finds that unified models generally underperform their base VLMs, and the Generate-then-Answer (GtA) inference often degrades performance compared to direct inference.

However, consistent improvements appear in spatial intelligence, visual illusions, and multi-round reasoning tasks.

Task-level and model-level correlations show that G2U gains cluster by cognitive demands and are strongly influenced by the base model's pretraining data, not just architecture.

The paper introduces two metrics, Reasoning-Alignment (RA) and Answer-Alignment (AL), to assess intermediate image quality.

The findings suggest that current unification paradigms may dilute general reasoning but offer promise for spatial and structural tasks, highlighting the need for better integration mechanisms and diverse training data.

02 From the paper

Abstract

Unified multimodal models have recently demonstrated strong generative capabilities, yet whether and when generation improves understanding remains unclear. Existing benchmarks lack a systematic exploration of the specific tasks where generation facilitates understanding. To this end, we introduce UniG2U-Bench, a comprehensive benchmark categorizing generation-to-understanding (G2U) evaluation into 7 regimes and 30 subtasks, requiring varying degrees of implicit or explicit visual transformations. Extensive evaluation of over 30 models reveals three core findings: 1) Unified models generally underperform their base Vision-Language Models (VLMs), and Generate-then-Answer (GtA) inference typically degrades performance relative to direct inference. 2) Consistent enhancements emerge in spatial intelligence, visual illusions, or multi-round reasoning subtasks, where enhanced spatial and shape perception, as well as multi-step intermediate image states, prove beneficial. 3) Tasks with similar reasoning structures and models sharing architectures exhibit correlated behaviors, suggesting that generation-understanding coupling induces class-consistent inductive biases over tasks, pretraining data, and model architectures. These findings highlight the necessity for more diverse training data and novel paradigms to fully unlock the potential of unified multimodal modeling.