Reproduce “Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality Assessment” on OpenResearch

About this paper

Shijie Zhao, Xuanyu Zhang, Weiqi Li, Junlin Li, Li Zhang, Tianfan Xue, Jian Zhang

This research analyzes how reinforcement learning (RL) enhances the generalization of Multimodal Large Language Models (MLLMs) in Image Quality Assessment (IQA) by compressing visual data into generalized text representations. The work introduces Reasoning-Aligned Cross-domain Training (RACT) for robust multi-dataset learning and Reasoning-Aligned Lightweight IQA (RALI), which achieves comparable performance to state-of-the-art reasoning MLLMs while reducing model parameters by over 95% and inference time/memory by over 95%.

Key insights

• RL training in MLLMs shifts dependency for IQA scoring from visual tokens to concise reasoning text, which acts as a compact (10x compression) and generalizable representation.
• Textual representations effectively mitigate domain discrepancies observed in visual feature spaces across different IQA datasets, enhancing out-of-domain generalization.
• The image-to-text reasoning process learned through RL is inherently highly consistent and generalizable across datasets, reducing overfitting risks even when reasoning capabilities are disabled during evaluation.

Summary by alphaXiv.