VersusQ: Pairwise Margin Reasoning for Generalizable Video Quality Assessment

2026-05-20

AI Summary

Existing video quality assessment methods rely on absolute scores, rendering them susceptible to annotation biases in datasets and limiting their cross-domain generalization. This work proposes VersusQ, a novel framework that introduces, for the first time, a purely relative comparison paradigm. Leveraging multimodal large language models, VersusQ performs fine-grained comparisons of video pairs in terms of both visual and temporal quality, producing signed continuous margin values that jointly encode preference direction and magnitude of difference. The framework employs a Margin-Coupled GRPO algorithm to jointly optimize relational reasoning and margin regression while generating interpretable justifications for each comparison. Experimental results demonstrate that VersusQ achieves state-of-the-art performance across multiple public benchmarks, significantly enhancing cross-domain generalization and ranking stability.

Abstract

Large Multimodal Models (LMMs) have shown promise for video quality assessment, but most methods still predict an absolute score for each video. Such pointwise supervision often mixes perceptual quality with dataset-specific calibration, including annotation protocols, rating habits, and score distributions. As a result, the learned scoring rule may work well within a benchmark but transfer poorly across unseen domains. We argue that relative comparisons alleviate the absolute-scale calibration bias by focusing purely on perceptual differences rather than dataset-specific rating habits. Consequently, we propose \textbf{VersusQ}, a pairwise margin reasoning framework driven entirely by direct comparisons. Specifically, VersusQ performs LMM-based comparison between two videos, reasons about their visual and temporal quality differences, and predicts a signed continuous margin that captures both the preferred choice and the degree of difference. Furthermore, to align interpretable comparison rationales with fine-grained numerical differences, we introduce Margin-Coupled GRPO, which jointly optimizes rollout-based relational reasoning and continuous margin regression. Extensive experiments on multiple public VQA benchmarks demonstrate that VersusQ achieves state-of-the-art performance, strong cross-domain generalization, and reliable fine-grained ranking under heterogeneous evaluation scenarios.

Problem

• video quality assessment
• absolute scoring bias
• cross-domain generalization
• dataset-specific calibration
• perceptual quality

Innovation

• pairwise comparison
• video quality assessment
• large multimodal models
• margin regression
• cross-domain generalization