Research map / Vision-language and multimodal / Multimodal LLMs

Refine-IQA: Multi-Stage Reinforcement Finetuning for Perceptual Image Quality Assessment

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 3 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 Refine-IQA, a two-stage RFT framework for LMM-based image quality assessment that first sharpens distortion perception on a new 20K-image dataset and then uses a probability-difference reward to supervise the "think" process during quality scoring.

Problem

Existing RFT-based IQA methods give no reward supervision to the "think" process, which collapses in length during GRPO training, and they fine-tune directly on IQA tasks without enhancing the model's native low-level distortion perception.

Approach

A multi-task rule-based reward for distortion type, severity, object and grounding (Stage-1), followed by a probability-difference reward that compares ground-truth MOS token likelihoods between think and no-think modes to implicitly supervise reasoning, plus token-level GRPO with a modified advantage that frames scoring as distribution prediction (Stage-2).

Key results

• Refine-IQA-S2 (w/o think) reaches 0.938 SRCC on KonIQ and 0.798 SRCC on AGIQA-3K, competitive with SFT-based LMMs on six IQA scoring datasets.
• Refine-IQA-S1 raises overall distortion-description accuracy on Refine-Perception-20K-test from 51.25% (Qwen2.5-VL-7B base) to 76.37%.
• Refine-IQA-S2 (w think) reaches 81.48% overall on Q-Bench-test, above GPT-4o (78.92%) and Qwen2.5-VL-7B w think (79.51%) on overall; Q-Instruct SFT reference reaches 81.52%.

Surprising findings

• Standard GRPO on the quality scoring task causes the think process to collapse in length while scoring accuracy keeps improving.
• Without the probability difference reward, the think-mode advantage over no-think mode nearly vanishes, and interpreting accuracy drops on Q-Bench.

Limitations

• Stage-1 uses synthetic distortions, so generalization to authentic distortions is not demonstrated.
• Reward design relies on a fixed MOS digit count and integer-part digit probabilities, which may limit generality.

Experimental setup

Datasets: Refine-Perception-20K KonIQ-10K SPAQ LSVQ Panda-70M COCO

Benchmarks: Refine-Perception-20K-test KonIQ SPAQ LIVE-C AGIQA-3K KADID-10K CSIQ Q-Bench-test

Metrics: SRCC PLCC accuracy IoU

Baselines: NIMA DBCNN HyperIQA MUSIQ CLIP-IQA+ LIQE Q-Align-IQA-7B Q-Align-OneAlign-7B Compare2Score-7B Q-Insight-7B Q-Instruct

Builds on: GRPO DAPO Q-Align Q-Insight

Closest papers by meaning

• VisualQuality-R1: Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rank (NeurIPS 2025 · similarity 0.90)
• Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality Assessment (ICLR 2026 · similarity 0.89)
• Q-Insight: Understanding Image Quality via Visual Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
• Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · similarity 0.89)
• Q-Probe: Scaling Image Quality Assessment to High Resolution via Context-Aware Agentic Probing (NeurIPS 2026 · similarity 0.88)
• QMamba: On First Exploration of Vision Mamba for Image Quality Assessment (ICML 2025 · similarity 0.88)