MR-IQA-2: Faithful Image Quality Reflection via Fine-Grained Credit Assignment

Yuan li; Youyuan Lin; Chenhui Chu; Shin'ya Nishida

cs.CVAug 19, 2026

MR-IQA-2: Faithful Image Quality Reflection via Fine-Grained Credit Assignment

Authors: Yuan li, Youyuan Lin, Chenhui Chu, Shin'ya Nishida

Organizations: Graduate School of Informatics, Kyoto University

Abstract

Multimodal large language models (MLLMs) have shown strong potential for image quality assessment (IQA) by improving consistency between quality ratings and their underlying reasoning. However, most approaches supervise reasoning through human-provided ratings and rarely examine whether it faithfully reflects image quality. Rating accuracy alone does not ensure faithful reasoning; a shared reward also obscures supervision sources and may reinforce unfaithful reasoning when a correct rating occurs by chance. To improve the faithfulness and reliability of blind IQA, we aim to (1) decouple credit assignment for reasoning and rating and (2) provide verifiable supervision for faithful reasoning. We introduce MR-IQA-2, an actor-editor-judge framework that operationalizes reasoning-editing-reflection. The actor generates quality reasoning for an input image, and the editor revises the image according to the identified quality factors. A frozen judge compares the original and edited images and provides reflective supervision for the actor's reasoning. MR-IQA-2 further uses fine-grained credit assignment to decouple reasoning and rating supervision. Judge feedback supervises reasoning, whereas human ratings supervise the predicted rating. Masked token-specific updates distinguish these signals while preserving the causal relation from reasoning to rating. Across IQA benchmarks, MR-IQA-2 achieves competitive rating alignment with humans. Visual reflection also enables richer and more faithful visual understanding beyond rating, which may inform image-quality optimization and related downstream tasks. Code is available at https://github.com/RobinY99/MR-IQA-2.

Figures & tables

Figure 1: Overview of MR-IQA-2. Prior MLLM-based IQA produces an explanation and a rating without visually verifying the explanation. Our actor converts a quality hypothesis into multiple targeted low-level edits. The independent judge compares the resulting quality changes and returns evidence-based feedback. Only visually supported interventions reinforce the actor, yielding faithful reasoning and a deeper understanding of causal quality-limiting factors. Here, $\Delta s$ denotes the Judge-score change between an edited image and the original image, as defined in Eq. ( 5 ).

Panel of Table 1 — Table 1: Rating performance comparison. Each dataset reports PLCC $\uparrow$ and SRCC $\uparrow$ between predicted and ground-truth ratings. Red and blue denote the best and second-best results, respectively. Baseline entries use reported results, except that VQ-R1 is reproduced with Qwen3-VL-2B ( Bai et al. 2025a ) because its original results were obtained under a different training protocol. Both MR-IQA-2 variants use Qwen3.5-4B; the unstarred row uses the final E5 credit-mask checkpoint, while MR-IQA-2 ∗ freezes the vision encoder and aligner during training. Zoom-IQA † uses additional annotations and combines SFT with RL; all other trainable methods use the same KonIQ training split.

Rating alignment
	Authentic						AI-generated		Synthetic				Average
	KonIQ		SPAQ		LIVE-W		AGIQA-3K		KADID-10k		CSIQ		Avg.
Method	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC	PLCC	SRCC
Baseline	0.553	0.583	0.750	0.743	0.592	0.564	0.679	0.655	0.660	0.669	0.681	0.657	0.652	0.645
Rating-only	0.952	0.938	0.901	0.899	0.896	0.875	0.808	0.740	0.669	0.676	0.789	0.764	0.836	0.815
Without credit mask (E5)	0.932	0.914	0.896	0.897	0.878	0.857	0.809	0.736	0.676	0.684	0.815	0.786	0.834	0.812

Table 3: Rating and quality-gain ablations. The upper panel reports PLCC/SRCC on the complete test sets and their six-dataset average. The lower panel reports Judge scores before and after reasoning-guided editing. Baseline and Rating-only use 5,843 common images; the E5 rows use their full successful audits (28,270 without the mask and 28,044 with the mask). Bold marks the best scalar result before rounding. The higher unmasked gain coincides with single-solution collapse and is not interpreted as improved reasoning.

Figure 2: Qualitative comparison before and after training. Each row shows the input image and the Actor’s reasoning-guided edits before and after training. Post-training reasoning yields more effective interventions and larger Judge-rated quality gains. Green text highlights additional quality factors identified after training.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Variant	Iter.	Time (h/ep.)	Val. P/S/M	U-score (%)	Gen. P/S/M
(a) Qwen3-VL-2B
T1: DAPO i1, frozen, global	1	1.38	0.831/0.817/0.868	12.0	0.809/0.787/0.936
(b) Qwen3.5-2B
T2: DAPO i1, frozen, global	1	1.13	0.890/0.881/0.691	36.0	0.835 /0.808/0.740
T3: DAPO i4, frozen, global	4	2.47	0.910/0.904/0.265	21.5	0.834/0.809/0.541
T4: GRPO i1, frozen	1	0.81	0.886/0.883/0.845	34.0	0.829/0.800/0.878

Appendix

Table 4: Actor-only ablations at their reported endpoints.

\mathrm{P}/\mathrm{S}/\mathrm{M}

denotes PLCC/SRCC/MAE; Val. is the aligned 200-image set, and Gen. is the six-dataset average. U-score measures distinct validation ratings, and Time is the mean wall-clock time per epoch. Bold marks the best alignment result among the controlled Qwen3.5-2B variants; higher PLCC/SRCC and lower MAE are better. All variants use E3.

Figure 3: Validation trajectories through E3 on the common 200-image set. T7 denotes the 4B local-six run. Curves are single fixed-seed runs, so no error bands are shown. Higher PLCC/SRCC and lower MAE are better, while U-score diagnoses rating diversity. Color, marker, and line style consistently identify T1–T7.

Figure 4: Quality gain versus visual change scale. The panels relate the Judge-observed quality gain $\Delta s$ to PSNR, MS-SSIM ( Wang et al. 2003 ) , and LPIPS ( Zhang et al. 2018 ) between the original and edited images. Colors denote datasets, curves show smoothed trends, and $\rho$ denotes Spearman correlation.

Figure 5: Quality scores versus low-level image features. Rows denote datasets and columns denote luminance, contrast, saturation, sharpness, colorfulness, and entropy. Purple and orange show human MOS and Judge scores for original images, while blue shows Judge scores for edited images. Curves show smoothed trends; $\rho$ and dCor denote Spearman and distance correlations ( Székely et al. 2007 ) .

Figure 6: Behavioral analysis of reasoning-guided editing. (a) Judge-rated quality gain versus human MOS across six benchmarks. (b) Quality gain versus changes in low-level image features. (c) Human MOS versus reference-based similarity metrics between the original and edited images. (d) Frequent reasoning tokens before and after training.

Method	KonIQ	SPAQ	LIVE-W	AGIQA-3K	KADID-10k	CSIQ	Average	MAE $\downarrow$
Without Mask (E1)	$\mathbf{0.952/0.938}$	$\mathbf{0.901}/0.898$	$0.897/0.877$	$\mathbf{0.816/0.747}$	$0.661/0.668$	$0.771/0.746$	$0.833/\mathbf{0.812}$	$\mathbf{0.492}$
Mask, no KL (E1)	$0.945/0.931$	$0.894/0.892$	$0.892/0.867$	$0.805/0.719$	$0.641/0.650$	$0.758/0.727$	$0.822/0.798$	$0.706$
Mask + KL (E1)	$0.951/0.938$	$0.896/0.895$	$\mathbf{0.903/0.880}$	$0.801/0.731$	$0.653/0.664$	$0.784/0.758$	$0.831/0.811$	$0.676$
Without Mask (E5)	$0.932/0.914$	$0.896/0.897$	$0.878/0.857$	$0.809/0.736$	$\mathbf{0.676/0.684}$	$0.815/\mathbf{0.786}$	$0.834/0.812$	$1.054$
Mask (E5)	$0.937/0.917$	$0.900/\mathbf{0.899}$	$0.893/0.863$	$0.809/0.739$	$0.667/0.669$	$\mathbf{0.824}/0.785$	$\mathbf{0.838}/0.812$	$0.543$

Appendix

Table 5: Dataset-level rating performance of the credit-routing variants. Dataset and Average entries report PLCC/SRCC, where Average is the unweighted average over the six datasets. The three E1 variants share the same source checkpoint; the two E5 runs are trained from the native parent for five epochs. The first two E1 variants use no KL. Bold marks the best result before rounding in each column.

	Rating	Functional reasoning		Language diversity
Method	$\mathrm{P}/\mathrm{S}$ $\uparrow$	Mean $\Delta s$ $\uparrow$	Pos. (%) $\uparrow$	Evidence unique (%) $\uparrow$	Evidence top-1 (%) $\downarrow$	Solution unique (%) $\uparrow$	Solution top-1 (%) $\downarrow$
Baseline	$0.652/0.645$	$-0.136$	$40.664$	–	–	–	–
Rating-only	$0.836/\mathbf{0.815}$	$-0.260$	$30.686$	–	–	–	–
Without Mask (E1)	$0.833/0.812$	$+0.837$	$96.954$	0.013	99.992	50.621	0.822
Mask, no KL (E1)	$0.822/0.798$	$+0.917$	$97.707$	1.233	21.196	94.508	0.085
Mask + KL (E1)	$0.831/0.811$	$+0.788$	$95.054$	24.018	13.191	75.228	0.191

Appendix

Table 6: Joint rating, functional reasoning, and diversity analysis.

\mathrm{P}/\mathrm{S}

is the six-dataset average PLCC/SRCC.

\Delta s

and Pos. are measured on 5,843 common images. Diversity columns report percentages; lower top-1 means less concentration. Unmarked diversity values use 23,599 matched inputs. † marks full E5 audits (28,270 Without Mask; 28,044 Mask).

Figure 7: Training dynamics at the retained checkpoints. Each point is the exact four-rank mean for E1, E2, E3, or E5. Without Mask produces larger reasoning reward and Judge gain, whereas rating reward converges similarly. Its concurrent reduction in completion length is consistent with convergence toward a high-reward solution template.

Figure 8: Checkpoint validation dynamics. Panels (a)–(c) report PLCC, SRCC, and Judge gain on the complete 200-image validation set. Panel (d) reports normalized-solution uniqueness. Without Mask maintains rating alignment and high Judge gain while its solutions collapse.

Explore similar work

Multi-Granularity Reasoning for Image Quality Assessment via Attribute-Aware Reinforcement Learning to Rank
Apr 7, 2026Xiangyong Chen, Xiaochuan Lin, Haoran Liu +3Image Quality Assessment VLM Reasoning
Henan Polytechnic University
IQA-T1: Tool-based Visual Evidence Reasoning for Image Quality Assessment
Jul 14, 2026Jinjian Wu, Jiaqi Tang, Wei Wei +5Tool Use in VLMs VLM Reasoning
School of Computer Science, Northwestern Polytechnical University, Xi’an, China · The Hong Kong University of Science and Technology, Hong Kong SAR, China
ME-IQA: Memory-Enhanced Image Quality Assessment via Re-Ranking
Mar 21, 2026Kanglong Fan, Tianhe Wu, Wen Wen +6Memory-Augmented VLMs Vision-Language Models
City University of Hong Kong, Hong Kong SAR, China · ByteDance Inc., China

Apr 7, 2026J/K move · Enter open · S save

Multi-Granularity Reasoning for Image Quality Assessment via Attribute-Aware Reinforcement Learning to Rank

Xiangyong Chen, Xiaochuan Lin, Haoran Liu, Xuan Li+2

Henan Polytechnic University

Image Quality Assessment VLM Reasoning RL for VLMs

Open paper AskPDF arXiv

Recent advances in reasoning-induced image quality assessment (IQA) have demonstrated the power of reinforcement learning to rank (RL2R) for training vision-language models (VLMs) to assess perceptual quality. However, existing approaches operate at a single granularity, predicting only an overall quality score, while overlooking the multi-dimensional nature of human quality perception, which encompasses attributes such as sharpness, color fidelity, noise level, and compositional aesthetics. In this paper, we propose MG-IQA (Multi-Granularity IQA), a multi-granularity reasoning framework that extends RL2R to jointly assess overall image quality and fine-grained quality attributes within a single inference pass. Our approach introduces three key innovations: (1) an attribute-aware prompting strategy that elicits structured multi-attribute reasoning from VLMs; (2) a multi-dimensional Thurstone reward model that computes attribute-specific fidelity rewards for group relative policy optimization; and (3) a cross-domain alignment mechanism that enables stable joint training across synthetic distortion, authentic distortion, and AI-generated image datasets without perceptual scale re-alignment. Extensive experiments on eight IQA benchmarks demonstrate that MG-IQA consistently outperforms state-of-the-art methods in both overall quality prediction (average SRCC improvement of 2.1%) and attribute-level assessment, while generating interpretable, human-aligned quality descriptions.

Abstract

Explore similar work

Multi-Granularity Reasoning for Image Quality Assessment via Attribute-Aware Reinforcement Learning to Rank

IQA-T1: Tool-based Visual Evidence Reasoning for Image Quality Assessment

ME-IQA: Memory-Enhanced Image Quality Assessment via Re-Ranking