Q-Probe: Scaling Image Quality Assessment to High Resolution via Context-Aware Agentic Probing

Q-Probe addresses a specific failure mode in modern MLLMs: the inability to detect fine-grained artifacts (noise, compression, blurring) in high-resolution images because they typically downsample inputs to a fixed resolution (e.g., 336x336). While 0 stars is typical for a brand-new arXiv release (8 days old), the 7 forks indicate immediate interest from the research community. Its defensibility is currently rooted in the 'agentic' logic that prevents the model from assuming a crop is automatically 'bad quality' (a common bias in IQA models). However, the moat is shallow; as frontier models (GPT-4o, Claude 3.5) move toward native high-res support or dynamic tiling, the need for a specialized probing agent diminishes. This project is highly valuable as a reference for niche IQA tasks (medical imaging, professional photography assessment) but faces risk from general-purpose vision improvements in 1-2 years. Competitors include generalist MLLMs and specialized IQA models like HyperIQA or MUSIQ, though Q-Probe's RL-driven agentic approach is a more 'modern' take on the problem.