Research map / Vision-language and multimodal / Multimodal LLMs

Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality Assessment

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning

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.

Shows that RL-trained reasoning IQA models generalize by compressing visual tokens into compact quality-reasoning text, then proposes RALI, a reasoning-free CLIP-based IQA model with bucketed k-means score bases that matches Q-Insight at ~4% of its parameters, plus RACT for cross-domain RL co-training.

Problem

Reasoning-based IQA models trained with visual RL generalize well but are slow and memory-heavy at inference; the mechanism behind their generalization is also unclear.

Approach

Interprets RL-trained MLLM reasoning as an image-to-text compression that yields domain-bridging representations, then distills this into a CLIP vision encoder aligned contrastively to reasoning text, with a PCA-reduced, bucketed k-means basis of scored prototypes so inference needs no LLM or reasoning.

Key results

• RALI reaches average PLCC/SRCC of 0.798/0.779 on single-dataset (KonIQ-trained) regression, close to Q-Insight's 0.806/0.783, while using about 4% of its parameters.
• RACT achieves out-of-domain average PLCC/SRCC of 0.858/0.816 on LiveW, AGIQA, CSIQ, TID2013 when co-trained on KonIQ, SPAQ, KADID, PIPAL, versus 0.796/0.740 for Q-Insight.
• At batch size 16 on an A100, RALI uses 14.7% of Q-Insight's memory and 3.4% of its inference time.

Surprising findings

• Adding score labels from multiple datasets to cross-domain SFT degraded out-of-domain performance, attributed to annotator bias.
• Q-Insight with reasoning disabled (GRPO without think) showed a large drop in out-of-domain performance, e.g., average PLCC 0.768 vs 0.806.

Limitations

• Performance ceiling bounded by the CLIP vision encoder.
• Experiments limited to natural-image IQA.

Experimental setup

Datasets: KonIQ SPAQ KADID PIPAL LIVE-Wild CSIQ TID2013 AGIQA

Benchmarks: KonIQ SPAQ KADID PIPAL LiveW AGIQA CSIQ TID2013

Metrics: PLCC SRCC

Baselines: Q-Insight VisualQuality-R1 Q-Align DeQA-Score CLIP-IQA+ LIQE MUSIQ ManIQA

Builds on: Q-Insight GRPO CLIP VisualQuality-R1 Q-Align DeQA-Score

Closest papers by meaning

• VisualQuality-R1: Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rank (NeurIPS 2025 · similarity 0.93)
• Q-Insight: Understanding Image Quality via Visual Reinforcement Learning (NeurIPS 2025 · similarity 0.92)
• Refine-IQA: Multi-Stage Reinforcement Finetuning for Perceptual Image Quality Assessment (AAAI 2026 · similarity 0.89)
• Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning of Vision Language Models (NeurIPS 2025 · similarity 0.89)
• On Robustness and Chain-of-Thought Consistency of RL-Finetuned VLMs (ICML 2026 · similarity 0.89)
• MiCo: Multi-image Contrast for Reinforcement Visual Reasoning (NeurIPS 2025 · similarity 0.89)