Research map / Vision-language and multimodal / Multimodal LLMs

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

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Q-Probe, an agentic zoom-in IQA MLLM trained with GRPO→SFT→GRPO on a new high-resolution Vista-Bench, using context-aware cropping to avoid the "cropping implies degradation" bias.

Problem

RL-based MLLM image quality assessment relies on global views and misses subtle local degradations in high-resolution images; naively adding zoom-in tools induces spurious cropping-implies-low-quality biases and misreads depth-of-field bokeh as artifacts.

Approach

An agentic IQA model that interleaves global perception with crop tool calls, trained through a three-stage curriculum (pairwise GRPO ranking on MOS, SFT on context-aware crop trajectories that include clean/bokeh regions, then decoupled RL with accuracy and localization rewards).

Key results

• Q-Probe reaches SRCC 0.801 on Vista-Bench versus 0.517 for VisualQuality-R1 and 0.450 for Qwen2.5-VL-7B.
• Average SRCC across seven benchmarks is 0.801 for Q-Probe versus 0.749 for VisualQuality-R1 (as reported in main table).
• Ablation: Stage 1 only gives 0.352 SRCC on Vista-Bench; full three-stage model gives 0.801, and crop strategy ablation drops to 0.578 when cropping only exact degradation regions.

Surprising findings

• Standard SFT as Stage 1 is reported to entrench global-view bias and hinder later zoom-in learning, motivating the GRPO-first curriculum.
• Restricting crops to exact degradation regions causes severe overfitting to a 'cropping implies low quality' bias (SRCC 0.578).

Limitations

• Vista-Bench is built by synthetic wavelet-based degradation injection on public high-res images, so artifacts may not match real-world distortions.
• Inserting zero-penalty background crops reduces localization precision, requiring a further RL stage.
• Average SRCC on some standard benchmarks (e.g., PIPAL) remains marginal over baselines.

Experimental setup

Datasets: KADID-10k HR-Bench 4K Pixel-Reasoner UHD-IQA KonIQ-10k SPAQ PIPAL TID2013 AGIQA-3K

Benchmarks: Vista-Bench SPAQ KADID-10k PIPAL TID2013 KonIQ-10k AGIQA-3K

Metrics: SRCC PLCC

Baselines: BRISQUE NIQE MUSIQ UNIQUE MANIQA Qwen2.5-VL-7B LIQE DeQA-Score Q-Align UnifiedReward-T Q-Insight VisualQuality-R1

Builds on: Q-Insight VisualQuality-R1 DeepEyes Pixel Reasoner Qwen2.5-VL

Closest papers by meaning

• Q-Insight: Understanding Image Quality via Visual Reinforcement Learning (NeurIPS 2025 · similarity 0.91)
• VisualQuality-R1: Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rank (NeurIPS 2025 · similarity 0.89)
• Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality Assessment (ICLR 2026 · similarity 0.89)
• Refine-IQA: Multi-Stage Reinforcement Finetuning for Perceptual Image Quality Assessment (AAAI 2026 · similarity 0.88)
• Self-Evolving Vision-Language Models for Image Quality Assessment via Voting and Ranking (ICLR 2026 · similarity 0.87)
• Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · similarity 0.87)