VQAThinker: Exploring Generalizable and Explainable Video Quality Assessment via Reinforcement Learning

TL;DR

VQAThinker introduces a reasoning-driven framework that combines large multimodal models with Group Relative Policy Optimization to address no-reference video quality assessment. By integrating three specialized rewards—bell-shaped regression, pairwise ranking, and temporal consistency—along with a format-guided reasoning trace, the approach achieves strong generalization to both in-domain and out-of-distribution videos and provides interpretable quality explanations. The method demonstrates state-of-the-art scoring and superior video quality understanding on tasks like distortion attribution and quality description, without requiring instruction-tuned data. This RL-based design offers a practical pathway to robust, explainable VQA applicable to diverse real-world video content and distortions.

Abstract

Video quality assessment (VQA) aims to objectively quantify perceptual quality degradation in alignment with human visual perception. Despite recent advances, existing VQA models still suffer from two critical limitations: \textit{poor generalization to out-of-distribution (OOD) videos} and \textit{limited explainability}, which restrict their applicability in real-world scenarios. To address these challenges, we propose \textbf{VQAThinker}, a reasoning-based VQA framework that leverages large multimodal models (LMMs) with reinforcement learning to jointly model video quality understanding and scoring, emulating human perceptual decision-making. Specifically, we adopt group relative policy optimization (GRPO), a rule-guided reinforcement learning algorithm that enables reasoning over video quality under score-level supervision, and introduce three VQA-specific rewards: (1) a \textbf{bell-shaped regression reward} that increases rapidly as the prediction error decreases and becomes progressively less sensitive near the ground truth; (2) a \textbf{pairwise ranking reward} that guides the model to correctly determine the relative quality between video pairs; and (3) a \textbf{temporal consistency reward} that encourages the model to prefer temporally coherent videos over their perturbed counterparts. Extensive experiments demonstrate that VQAThinker achieves state-of-the-art performance on both in-domain and OOD VQA benchmarks, showing strong generalization for video quality scoring. Furthermore, evaluations on video quality understanding tasks validate its superiority in distortion attribution and quality description compared to existing explainable VQA models and LMMs. These findings demonstrate that reinforcement learning offers an effective pathway toward building generalizable and explainable VQA models solely with score-level supervision.

Figures (5)

• Figure 1: Comparison of Q-Insight, VisualQuality-R1, and our VQAThinker in video quality understanding and scoring. All three models are trained on the LSVQ dataset. Compared to Q-Insight and VisualQuality-R1, VQAThinker generates a more comprehensive reasoning trace that covers major distortion types, thereby producing more accurate video quality scores.
• Figure 2: Overall framework of VQAThinker. The architecture consists of a LMM equipped with a frozen motion encoder and a motion projector. During training, the raw video is degraded using a frame perturbation operator to generate a perturbed version. Regression and ranking rewards are computed for both the raw and perturbed videos, while the temporal reward is obtained by comparing the reward differences between them. The regression, ranking, and temporal rewards derived from the raw videos are then used to optimize VQAThinker via the GRPO algorithm.
• Figure 3: The structure of our LMM.
• Figure 4: Illustration of different temporal perturbation modes.
• Figure 5: Examples illustrating our model's visual quality scoring and understanding capability. All MOS scores are linearly rescaled to a unified range of 1–5.