Research map / Vision-language and multimodal / Multimodal LLMs

Self-Evolving Vision-Language Models for Image Quality Assessment via Voting and Ranking

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis

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 EvoQuality, a label-free self-evolving framework where a VLM creates pairwise pseudo-rankings by majority voting over its own outputs and is trained with GRPO using a fidelity reward for image quality assessment.

Problem

Post-training VLMs for image quality assessment usually needs costly human-annotated quality labels; self-consistency-style self-supervision has not been explored for perceptual IQA, where quality is relative and subjective.

Approach

Adapts self-consistency to ranking-based IQA: pairwise majority voting on the VLM's own comparisons yields pseudo-preference labels, which are converted into a Thurstone-style fidelity reward and optimized with GRPO in iterative offline/online rounds.

Key results

• EvoQuality@Round2 raises weighted-average PLCC from 0.615 to 0.770 (+31.8%) over the Qwen2.5-VL-7B baseline on zero-shot IQA benchmarks.
• EvoQuality@Round2 reaches weighted-average SRCC 0.719 versus Q-Insight's 0.667 among supervised VLM-based models trained on KONIQ (per the paper's Table 2 comparison).
• Stacking EvoQuality on pre-trained VisualQuality-R1 raises weighted-average PLCC from 0.722 to 0.768 in the co-training setting.

Surprising findings

• Direct score regression with voted pseudo-scores (EvoEstimate) plateaus in round 2, while the pairwise ranking variant keeps improving.
• EvoQuality trails supervised VLM models on in-distribution KONIQ (0.835 vs 0.953 PLCC for DeQA-Score).

Limitations

• Depends on an in-domain pool of unlabeled images; diversity of that corpus limits performance.
• Self-evolution with iterative voting and RL incurs significant computational overhead.
• Lags supervised VLM methods on in-distribution KONIQ performance.

Experimental setup

Datasets: KONIQ SPAQ KADID

Benchmarks: KONIQ SPAQ AGIQA LIVEW KADID PIPAL TID2013 CSIQ

Metrics: PLCC SRCC weighted average

Baselines: NIQE BRISQUE NIMA HyperIQA DBCNN CONTRIQUE Re-IQA MUSIQ CLIP-IQA+ MANIQA Q-Align DeQA-Score

Builds on: GRPO TTRL VisualQuality-R1 Thurstone model Q-Insight

Closest papers by meaning

• VisualQuality-R1: Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rank (NeurIPS 2025 · similarity 0.91)
• Q-Insight: Understanding Image Quality via Visual Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
• Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality Assessment (ICLR 2026 · similarity 0.88)
• Revisiting MLLM Based Image Quality Assessment: Errors and Remedy (AAAI 2026 · similarity 0.87)
• Scaling Test-Time Robustness of Vision-Language Models via Self-Critical Inference Framework (CVPR 2026 · similarity 0.87)
• Q-Probe: Scaling Image Quality Assessment to High Resolution via Context-Aware Agentic Probing (NeurIPS 2026 · similarity 0.87)