UNO-Bench: A Unified Benchmark for Exploring the Compositional Law Between Uni-modal and Omni-modal in Omni Models

TL;DR

UNO-Bench introduces a unified Omni-model benchmark that simultaneously evaluates uni-modal and omni-modal capabilities across 44 task types and 5 modality combinations. It couples a human-curated omni-modal dataset with enhanced uni-modal data, a novel Multi-Step Open-Ended Question format, and a general scoring model to enable automated, scalable evaluation. The core finding is a multiplicative Compositional Law, $P_{Omni} = C \cdot (P_{A} \cdot P_{V})^{\alpha} + b$, with $\alpha \approx 2.19$, $C \approx 1.03$, and $b \approx 0.24$, indicating strong power-law synergy and emergent omni-modal capabilities as uni-modal foundations improve. Experiments reveal omni-modal bottlenecks on weaker models but pronounced synergy on stronger models, with Gemini-2.5-Pro approaching human-like perception yet lagging in reasoning, highlighting the need to advance cross-modal reasoning. The benchmark also demonstrates high data quality and evaluation efficiency via clustering-based compression (CGHS) and supports future expansion to harder tasks and more modalities.

Abstract

Multimodal Large Languages models have been progressing from uni-modal understanding toward unifying visual, audio and language modalities, collectively termed omni models. However, the correlation between uni-modal and omni-modal remains unclear, which requires comprehensive evaluation to drive omni model's intelligence evolution. In this work, we introduce a novel, high-quality, and UNified Omni model benchmark, UNO-Bench. This benchmark is designed to effectively evaluate both UNi-modal and Omni-modal capabilities under a unified ability taxonomy, spanning 44 task types and 5 modality combinations. It includes 1250 human curated samples for omni-modal with 98% cross-modality solvability, and 2480 enhanced uni-modal samples. The human-generated dataset is well-suited to real-world scenarios, particularly within the Chinese context, whereas the automatically compressed dataset offers a 90% increase in speed and maintains 98% consistency across 18 public benchmarks. In addition to traditional multi-choice questions, we propose an innovative multi-step open-ended question format to assess complex reasoning. A general scoring model is incorporated, supporting 6 question types for automated evaluation with 95% accuracy. Experimental result shows the Compositional Law between omni-modal and uni-modal performance and the omni-modal capability manifests as a bottleneck effect on weak models, while exhibiting synergistic promotion on strong models.

Figures (18)

• Figure 1: Benchmark Statistics and Evaluation Results.
• Figure 2: Illustration of the unified ability taxonomy proposed in UNO-Bench.
• Figure 3: Illustration of the cross-modality solvable sample.
• Figure 4: Dataset Construction Pipeline includes human-centric process (left side) and automated data compression (right side). First, we collect diverse and novel materials to prevent data contamination. Second, with the proposed unified ability taxonomy, human annotators including experts will craft questions, answers and record audios in real-world scenarios. Finally, with model checking, ablation study and human experts revision, we achieves high quality and diversity dataset. Regarding automated data compression, we present a clustering-guided hierarchical sampling method to achieve efficient compression while maintaining high evaluation consistency.
• Figure 5: Construction of Multi-Step Open-Ended Questions.