UniG2U-Bench: Do Unified Models Advance Multimodal Understanding?

TL;DR

This paper introduces UniG2U-Bench, a comprehensive benchmark to evaluate whether generation capabilities in unified multimodal models enhance understanding, revealing that current models often underperform and that specific tasks benefit from improved spatial and reasoning abilities.

Contribution

The paper presents UniG2U-Bench, a new benchmark categorizing generation-to-understanding evaluation into 7 regimes and 30 subtasks, and provides extensive analysis of over 30 models.

Findings

Unified models often underperform compared to base VLMs.

Generation inference can degrade performance relative to direct inference.

Spatial reasoning and multi-step tasks benefit from enhanced perception and reasoning.

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.