Q-Router: Agentic Video Quality Assessment with Expert Model Routing

Shuo Xing, Soumik Dey, Mingyang Wu, Ashirbad Mishra, Naveen Ravipati, Binbin Li, Hansi Wu, Zhengzhong TuSubmission to ICLR 2026 · Rejected

TL;DR

Q-Router is an agentic framework for universal Video Quality Assessment (VQA) that uses a Vision-Language Model (VLM) as an adaptive orchestrator to dynamically select and fuse specialized expert models (e.g., UGC and AIGC evaluators). By coupling a budget-aware, multi-tier routing architecture with a spatio-temporal artifact localization pipeline, Q-Router delivers state-of-the-art accuracy, robust cross-domain generalization, and interpretable visual evidence of video distortions.

What Problem Does This Paper Address?

Traditional deep learning models for video quality assessment directly regress a single scalar score from raw pixels. While effective within their specific training domains, they suffer from three major shortcomings:

  1. Poor Cross-Domain Generalization: Video formats have expanded from user-generated content (UGC) to AI-generated content (AIGC). Standard UGC-trained evaluators degrade sharply when facing AIGC artifacts (e.g., a ~70% drop in PLCC for COVER).
  2. Limited Interpretability: Scalar quality ratings function as black boxes, failing to indicate why or where quality degrades.
  3. Inflexibility: Incorporating new distortion types or content classes typically requires complete retraining or architectural redesigns.

Methods

Q-Router addresses these challenges via an agentic, multi-tiered framework:

[Input Video] 
      │
      ▼
[VLM Router (e.g., GPT-4o)] ──> Classifies Domain & Allocates Weights
      │
      ├──> Tier 0: Single-Expert Fast Path (Coarse inference)
      ├──> Tier 1: Multi-Expert Weighted Fusion (Dynamic weighting)
      └──> Tier 2: Fine-Grained Artifact Localization Pipeline
  • Agentic Routing & Weighted Fusion: A VLM acts as an orchestrator. It evaluates sampled frames and metadata to classify the video (UGC, AIGC, or Computer Graphics), queries specialized expert models from an extensible pool (including COVER, DOVER, UVQ, MaxVQA, ModularBVQA, and T2VQA), and dynamically computes fused scores based on model-task fit and inter-model agreement.
  • Three-Tier Compute Hierarchy:
  • Tier 0: Routes to a single best-fit expert for minimal latency.
  • Tier 1: Blends predictions across multiple experts using relevance-aware dynamic weighting.
  • Tier 2: Augments multi-expert fusion with explicit spatio-temporal artifact localization.
  • Spatio-Temporal Artifact Localization (Tier 2):
  • Probabilistic Extraction: Identifies suspicious candidate frames using handcrafted indicators (motion residuals, sharpness energy, gradient kurtosis, edge density, and HSV histogram shifts) grouped by hysteresis thresholding.
  • VLM Semantic Filtering: GPT-4o filters false alarms, keeping only frames exhibiting hallucinations, image artifacts, or AI inconsistencies.
  • Perceptual Heatmap Generation: Successive frames are motion-compensated via optical flow, and spatial LPIPS calculates pixel-level difference maps highlighting local distortions.

Datasets

The authors evaluated Q-Router across diverse benchmarks:

  • UGC VQA Benchmarks: KoNViD-1k, LSVQ-1080, LSVQ-Test, LIVE-VQC, and YouTube-UGC (YT-UGC).
  • AIGC VQA Benchmark: T2VQA-DB (focused on text-to-video distortions).
  • Video Quality Question Answering: Q-Bench-Video (both development and test splits).

Key Insights and Ideas

  • Expert Ensembling Solves the UGC–AIGC Divide: Standalone models fail on out-of-domain distributions (e.g., DOVER achieves only 0.0642 PLCC on AIGC, while the AIGC model T2VQA manages just 0.1717 PLCC on KoNViD-1k). Routing by content semantics bridges this gap without retraining underlying models.
  • System 1 vs. System 2 Reasoning in VLM Scoring: Direct zero-shot quality prompting of VLMs correlated surprisingly well with human Mean Opinion Scores (MOS), whereas forced Chain-of-Thought (CoT) reasoning degraded performance—suggesting deliberative text reasoning can misalign with human perceptual intuition. Incorporating specialized expert predictions restored and surpassed this baseline.
  • Modular Test-Time Scaling: Instead of end-to-end black-box diagnostic models, combining classical motion-compensated perceptual features with VLM filtering creates reliable, model-agnostic localization heatmaps.

Key Results

  • Overall VQA Performance: Across both UGC and AIGC benchmarks, Q-Router (Tier 1) achieved the highest overall average performance (0.8539 PLCC / 0.8407 SRCC), outperforming all standalone baselines by over 20%.
  • AIGC Robustness: On T2VQA-DB, Q-Router attained 0.8283 PLCC, drastically exceeding conventional UGC models (PLCC < 0.25).
  • Video QA SOTA: On the Q-Bench-Video benchmark:
  • Dev Split: Tier 2 achieved 60.07% overall accuracy, outperforming proprietary GPT-4o (56.91%).
  • Test Split: Tier 1 scored 59.45% vs. GPT-4o’s 56.50%, with particularly large margins on AIGC-related questions (64.42% vs. 53.80%).
  • Model Agnostic: Ablation with an open-source 4B model (Qwen-VL) maintained competitive performance and a steady upward trend from Tier 0 (50.06%) to Tier 2 (51.42%), verifying the architecture's generalizability.

Limitations and Future Work

  • Inference Latency: VLM routing and optical-flow preprocessing introduce substantial computational overhead (routing latency ranges from 7.11s in Tier 0 to 16.36s in Tier 2 per video, excluding expert execution time), making higher tiers unsuited for real-time video streaming.
  • Dataset-Specific Peak Performance: On certain isolated benchmarks, the top specialized expert still slightly edges out the fused model (e.g., UVQ on YT-UGC), as Q-Router optimizes for cross-domain stability rather than single-dataset tuning.
  • Future Directions: Developing fully end-to-end, learned diagnostic localization modules to replace handcrafted feature stages, and investigating Q-Router's localized heatmaps as reward functions for post-training video generation models.

Summary updated

Abstract

Video quality assessment (VQA) is a fundamental computer vision task that aims to predict the perceptual quality of a given video in alignment with human judgments. Existing performant VQA models trained with direct score supervision suffer from **(1)** *poor generalization* across diverse content and tasks, ranging from user-generated content (UGC), short-form videos, to AI-generated content (AIGC), **(2)** *limited interpretability*, and **(3)** *lack of extensibility* to novel use cases or content types. We propose Q-Router, an agentic framework for universal VQA with a multi-tier model routing system. Q-Router integrates a diverse set of expert models and employs vision–language models (VLMs) as real-time routers that dynamically reason then ensemble the most appropriate experts conditioned on the input video semantics. We build a multi-tiered routing system based on the computing budget, with the heaviest tier involving a specific spatiotemporal artifacts localization for interpretability. This agentic design enables Q-Router to combine the complementary strengths of specialized experts, achieving both flexibility and robustness in delivering consistent performance across heterogeneous video sources and tasks. Extensive experiments demonstrate that Q-Router matches or surpasses state-of-the-art VQA models on a variety of benchmarks, while substantially improving generalization and interpretability. Moreover, Q-Router excels on the quality-based question answering benchmark, Q-Bench-Video, highlighting its promise as a foundation for next-generation VQA systems. Finally, we show that Q-Router can localize spatio-temporal artifact/hallucination localization, showing potential as a reward function for post-training video generation models.
BibTeX
@misc{xing2025q,
  author = {Shuo Xing and Soumik Dey and Mingyang Wu and Ashirbad Mishra and Naveen Ravipati and Binbin Li and Hansi Wu and Zhengzhong Tu},
  title = {{Q-Router: Agentic Video Quality Assessment with Expert Model Routing}},
  year = {2025},
  eprint = {2510.08789},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.08789},
}
Figure