Paper deep dive
MR-IQA-2: Faithful Image Quality Reflection via Fine-Grained Credit Assignment
Yuan li, Youyuan Lin, Chenhui Chu, Shin'ya Nishida
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 8/20/2026, 4:45:40 AM
Summary
The paper introduces MR-IQA-2, a framework for faithful Image Quality Assessment (IQA) using Multimodal Large Language Models (MLLMs). It addresses the issue of unfaithful reasoning by decoupling credit assignment for reasoning and rating via fine-grained masked token updates. The framework employs an actor-editor-judge architecture where an actor generates reasoning and ratings, an editor modifies the image based on the reasoning, and a judge evaluates the quality change to provide verifiable supervision for the reasoning process.
Entities (9)
Relation Signals (8)
MR-IQA-2 → evaluatedon → KonIQ-10k
confidence 95% · We train all controlled models on the training subset of the KonIQ-10k split... We evaluate on the KonIQ-10k test split
MR-IQA-2 → usescomponent → Editor
confidence 95% · MR-IQA-2, an actor-editor-judge framework... the editor revises the image
MR-IQA-2 → usescomponent → Judge
confidence 95% · MR-IQA-2, an actor-editor-judge framework... A frozen judge compares the original and edited images
MR-IQA-2 → usescomponent → Actor
confidence 95% · MR-IQA-2, an actor-editor-judge framework... The actor generates quality reasoning
MR-IQA-2 → usesmethod → Fine-Grained Credit Assignment
confidence 95% · MR-IQA-2 further uses fine-grained credit assignment to decouple reasoning and rating supervision.
Editor → initializedwith → FLUX.2-klein
confidence 90% · The Editor uses FLUX.2 [klein] 4B
Actor → initializedwith → Qwen3.5-4b
confidence 90% · The Actor is initialized from Qwen3.5-4B
MR-IQA-2 → optimizeswith → GRPO
confidence 90% · We optimize the Actor with Group Relative Policy Optimization (GRPO)
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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 this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2608.18579v1
- Canonical: https://arxiv.org/abs/2608.18579v1
Trouble viewing inline? Open PDF directly →
Full Text
71,123 characters extracted from source content.
Expand or collapse full text
MR-IQA-2: Faithful Image Quality Reflection via Fine-Grained Credit Assignment Yuan Li, Youyuan Lin, Chenhui Chu, Shin’ya Nishida Graduate School of Informatics, Kyoto University MR-IQA-2 Abstract Multimodal large language models (MLLMs) have shown great potential to improve image quality assessment (IQA) by making predictions more explainable through increased con- sistency between quality ratings and their underlying reason- ing. However, most existing approaches supervise reasoning only to align it with human-provided quality ratings, paying little attention to whether the reasoning faithfully reflects the image’s actual quality content. Since higher rating accuracy alone does not guarantee faithful reasoning, using a shared reward for both rating and reasoning obscures the source of supervision and may reinforce unfaithful reasoning when a correct rating is obtained by chance. To improve the faithful- ness and reliability of blind IQA, we propose to (1) decouple credit assignment for reasoning and rating, and (2) provide ver- ifiable supervision signals for faithful reasoning. To this end, we introduce MR-IQA-2, an actor-editor-judge framework for faithful IQA that operationalizes reasoning-editing-reflection. The actor first generates quality reasoning for an input image. Conditioned on this reasoning, the editor revises the image to provide a verifiable visual signal for the identified qual- ity factors. A frozen judge then compares the original and edited images, producing reflective supervision that improves the actor’s quality reasoning. MR-IQA-2 uses fine-grained credit assignment to decouple reasoning and rating supervi- sion. 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 hu- mans. Additionally, visual reflection enables the framework to acquire richer and more faithful visual understanding beyond rating. This faithful visual understanding may inform image- quality optimization and related downstream tasks. Code is available at MR-IQA-2. 1 Introduction Image quality assessment (IQA) aims to understand how humans perceive visual quality and to model the relation- ship between an image and its perceived quality rating. Prior IQA work has undergone a series of framework transitions as visual content and degradation patterns have become in- creasingly complex. The goal of IQA is no longer limited to producing a numerical rating. A more important question is whether a model can develop a faithful understanding of image quality. Such understanding should go beyond specific degradations and cover both low-level visual attributes and semantic-level perception. Early blind image quality assessment (BIQA) methods were largely driven by degradation-centered assumptions. In this setting, image quality was often associated with the strength of synthetic distortions, and the learning objec- tive was commonly formulated as degradation-scale estima- tion. Methods such as ARNIQA (Agnolucci et al. 2024), TOPIQ (Chen et al. 2024), and LIQE (Zhang et al. 2023) represent important attempts to learn quality-aware repre- sentations under this paradigm. With the emergence of practical IQA benchmarks such as KonIQ-10k (Hosu et al. 2020) and SPAQ (Fang et al. 2020), the focus of BIQA has gradually shifted. Real-world images contain diverse content, complex capture conditions, aesthetic factors, and semantic preferences. These factors cannot be fully described by synthetic distortion types. As a result, BIQA has moved from estimating degradation inten- sity toward understanding the image itself. Recent multimodal approaches further extend this trend. Works such as Q-Instruct (Wu et al. 2024), Q-Insight (Li et al. 2025a), and VisualQuality-R1 (Wu et al. 2025) introduce lan- guage, criteria, and human-like reasoning into BIQA. These methods allow models to evaluate images through quality dimensions, ranking preferences, and explanatory outputs. This transition improves interpretability and generalization via criteria closer to human perception. But it also raises a deeper question: does the generated reasoning faithfully reflect the actual factors that determine image quality? This question is critical for BIQA based on multimodal large language models (MLLMs). MLLMs can produce flu- ent and plausible quality explanations, but plausible language does not guarantee faithful visual reasoning. A model may attribute low quality to reasonable-sounding factors while failing to identify the true visual limitations. Causal reason- ing collapse, unverifiable explanations, and deviations from human perception can therefore appear in quality assessment. Previous studies, including BRIQA (Li et al. 2025b) and H- IQA (Li et al. 2025c), have attempted to analyze reasoning failures, improve reasoning consistency, and align IQA with human perception. However, the reasoning process itself re- mains difficult to verify. Recent methods, including Zoom-IQA (Liang et al. 2026) and Tool-IQA (Qin et al. 2026), move BIQA toward arXiv:2608.18579v1 [cs.CV] 19 Aug 2026 tool-augmented reasoning through region-aware inspection. These methods show that additional visual observations help models inspect image details more reliably. Nevertheless, they focus on observation and grounding rather than verify- ing whether the proposed factors actually limit image quality. It therefore remains unclear whether a model has identified the factors that genuinely constrain overall image quality. Another persistent issue in reinforcement learning (RL)- based BIQA methods, including Q-Insight (Li et al. 2025a), VisualQuality-R1 (Wu et al. 2025), MR-IQA (Li et al. 2026), Zoom-IQA (Liang et al. 2026), and Tool-IQA (Qin et al. 2026), is that reasoning supervision is often determined by rating performance. Consequently, incorrect reasoning may still receive high rewards when the rating is accurate, poten- tially undermining reasoning faithfulness. In this work, we decompose the learning of faithful qual- ity reasoning into two tasks: (1) decoupling reasoning and rating supervision and (2) constructing an independent su- pervision signal for reasoning. For the first task, we pro- pose fine-grained credit assignment. We use masked token- specific updates so that better reasoning or a more accurate rating is rewarded accordingly within its corresponding out- put while preserving causality. The second task is therefore to construct a reliable supervision signal for reasoning. We use visual reflection to estimate this signal from observed quality changes. A reasoning proposal receives higher credit when its corresponding edit alleviates the identified quality- limiting factors. Finally, we integrate these two mechanisms into an actor-editor-judge framework termed MR-IQA-2. Prior work primarily pursues stronger rating perfor- mance, as exemplified by MR-IQA (Li et al. 2026), or richer reasoning processes, as in Zoom-IQA (Liang et al. 2026) and Tool-IQA (Qin et al. 2026). In contrast, MR- IQA-2 investigates whether the generated reasoning faith- fully identifies quality-limiting factors and yields visual understanding that can inform broader downstream tasks, including image editing and visual recognition. The main contributions are summarized as follows: • We introduce fine-grained credit assignment for the reasoning–rating structure in BIQA. Masked token- specific updates decouple reasoning and rating supervi- sion, assigning signal to its corresponding output while preserving the causal relation from reasoning to rating. • We construct a verifiable supervision signal for faithful quality reasoning through visual reflection. Reasoning- conditioned editing serves as a visual intervention, while a frozen judge evaluates the observed quality change and assigns credit to the corresponding reasoning. • We integrate these two mechanisms into MR-IQA-2, a modular actor-editor-judge framework that operational- izes reasoning-editing-reflection. The actor, editor, and judge can be independently instantiated or replaced for different models and downstream settings. 2 Related Work 2.1 Faithful Quality Reasoning with MLLMs. The adoption of MLLMs has improved the interpretability of BIQA by enabling language-based image understanding and quality explanations. However, faithful quality reasoning re- mains a key challenge. Early methods such as DepictQA (You et al. 2024) and Q-Instruct (Wu et al. 2024) mainly rely on su- pervised fine-tuning (SFT), which encourages plausible ex- planation patterns. Constrained by their training data, these models may overfit templated responses and exhibit limited logical coherence. Later RL-based frameworks, including Q- Insight (Li et al. 2025a), VisualQuality-R1 (Wu et al. 2025), H-IQA (Li et al. 2025c), and MR-IQA (Li et al. 2026), optimize generated reasoning using rating supervision. By strengthening the model’s underlying perceptual capabili- ties, these methods mitigate template overfitting, produce richer and more consistent reasoning, and improve rating alignment. Nevertheless, the reasoning itself is generated by the MLLM backbone without a reliable reasoning-specific supervision signal, making its faithfulness difficult to ver- ify. Our framework instead evaluates reasoning faithfulness through visual intervention by verifying whether edits de- rived from the identified reasoning factors lead to observable improvements in image quality. 2.2 Tool-Augmented IQA Methods Before MLLM-based BIQA, degradation modeling and task- specific image processing were already used to derive quality cues. ARNIQA (Agnolucci et al. 2024) applies predefined degradation operators to construct contrastive views and learn distortion-aware representations. More recently, Zoom- IQA (Liang et al. 2026) introduces region-aware zooming for detail inspection. Tool-IQA (Qin et al. 2026) further employs a magnifier and a gamma corrector to provide additional vi- sual evidence for quality rating. Although these tools enable richer and more localized inspection, the resulting reasoning is still evaluated mainly through rating performance. Rather than enriching the visual evidence during reasoning for rat- ing optimization, our design provides direct supervision for the faithfulness of quality reasoning. 3 Methods 3.1 MR-IQA-2 Framework Overview In this work, we propose MR-IQA-2 (Figure 1), an actor- editor-judge framework for faithful IQA. MR-IQA-2 makes quality reasoning verifiable rather than assessing it solely through rating performance. The actor is a general MLLM- based BIQA model that produces interpretable quality rea- soning and a rating. A frozen editor, initialized from an image-editing diffusion model, applies edits conditioned on the reasoning. A frozen MLLM judge then evaluates the quality change between the original and edited images. The observed change provides a reasoning-specific supervision signal for the actor. The following sections introduce (1) the interaction among the actor, editor, and judge, (2) fine- grained credit assignment, and (3) the training optimization procedure. 3.2 Actor-Editor-Judge Trajectory This section describes the interaction flow among the Ac- tor, Editor, and Judge. For a local batch of N images, LLM Reasoning Tokens: Evidence, Solution Rating Tokens Actor Assess the overall perceptual quality. Respond with exactly containing: "reasoning" and "rating". "reasoning" contains: "evidence" and "solution". “Evidence“: visible image evidence that determines its current overall perceptual quality and location. “Solution“: a coherent image-edit plan that causally addresses the evidence. Retain the same semantic meaning as the input. “Rating“: a numeric string from 1.00 to 5.00. "1.00" represents the worst quality and "5.00" represents excellent quality. Prompt Input Evidence: Although lighting is uneven, the key features and painted designs are discernible and draw the attention. Solution: Correct the slightly overexposed highlights on the vase to restore natural, balanced illumination. Rating:3.67 Evidence: Distracting purple glare in the lower-left area. Some highlights on the vase are slightly overexposed. Solution: Remove the distracting purple glare and color cast from the lower-left area, and correct the slightly overexposed highlights on the vase to restore illumination. Rating:3.40 Evidence: The purple lighting cast on the left side is an artifact of the environment or lighting setup, further degrading the image. Solution: Remove the unnatural purple lighting cast and glare from the lower-left area, restoring neutral illumination. Rating:2.52 K Samples Image Input Aligner Vision Encoder Vision Encoder Editing Model Editor Text Encoder Judge Masked Credit Assignment Original ImageReasoning TokensEdited Image 1Edited Image kEdited Image K Rollout 1 Rollout k Rollout K Reasoning: The vase is clearly recognizable, but strong purple lens flare, overexposed highlights, moderate softness. Rating:3.20 Judge Original Reasoning: Although lighting is uneven, the painted designs are discernible and draw the attention. Rating:3.30 Judge 1: ∆ s =0.10 Reasoning:The subject is clear and reasonably sharp, but strong highlights on the vase are slightly overexposed. Rating:3.70 Judge k: ∆ s =0.50 Reasoning:The vase is clear and detailed, though uneven lighting and slight softness reduce the overall quality. Rating:4.25 Judge K: ∆ s =1.05 Pretrained MLLM Good rating. Good reasoning. Reasoning: greater quality gains, higher credit. Optimization Target: (Rollout K’s Reasoning, Rollout k’s rating) Rating: better margin estimation, higher credit. Original and Edited Images trainable frozen masked token 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, ∆s denotes the Judge-score change between an edited image and the original image, as defined in Eq. (5). i ∈ 1,...,N indexes an image, K denotes the num- ber of Actor samples per image, and k ∈ 1,...,K in- dexes one sample. Within each training trajectory, super- script b∈0, 1 denotes the original-image and post-editing stages, respectively. Actor receives an original image I 0 i and uses a prompt P A to generate quality reasoning and a rating: a 0 i,k = r 0 i,k ,q 0 i,k ∼ π old ·| I 0 i ,P A ,(1) whereπ old denotes the original Actor policy used to generate the rollouts, and r 0 i,k and q 0 i,k denote reasoning and rating. Editor corrects the quality factors identified by the Actor: I 1 i,k = E φ E I 0 i ,r 0 i,k ;P E , φ E fixed.(2) where P E is a fixed editing prompt template and I 1 i,k is the edited output. Judge independently rates the original and edited images: s 0 i = J φ J I 0 i , s 1 i,k = J φ J I 1 i,k .(3) The resulting trajectory is T i,k = I 0 i ,a 0 i,k ,I 1 i,k ,s 0 i ,s 1 i,k .(4) The Judge-observed quality change is ∆s i,k = s 1 i,k − s 0 i .(5) 3.3 Fine-Grained Credit Assignment Motivation. As discussed in the Introduction, higher rat- ing accuracy alone does not guarantee faithful reasoning. Moreover, assigning a shared reward to reasoning and rating obscures the source of supervision. We therefore define sep- arate rewards for reasoning, rating, and format, and use each reward to update its corresponding output tokens. Reasoning reward. For each image, the Actor samples K reasoning–rating outputs indexed by k. Each reasoning pro- posal is evaluated with the frozen Editor and Judge. The editing prompt restricts interventions to low-level quality at- tributes while preserving image content. We measure reason- ing faithfulness by the resulting Judge-score improvement: R reasoning i,k = ∆s i,k .(6) Under these controlled conditions, reasoning that produces a larger quality improvement receives a higher reward. Rating reward. Following MR-IQA (Li et al. 2026), we supervise rating through relative quality margins. Let y i be the human mean opinion score (MOS). For another image j ̸= i in the same batch, the scale-controlled margin error is z rating i,k,j = q 0 i,k − 1 K P K k ′ =1 q 0 j,k ′ − (y i − y j ) τ ij , (7) where τ ij > 0 controls the margin-error scale. Using the L2 estimator, we average the N − 1 pairwise rewards: R rating i,k = 1 N − 1 N X j=1 j̸=i e − 1 2 ( z rating i,k,j ) 2 .(8) Format reward. The Actor output must be a valid JSON object containing the ordered fields reasoning and rating. LetF denote this output contract. The format reward is R format i,k = 1 a 0 i,k ∈F .(9) Masked credit assignment. Letc index a supervision type. For token t, χ c i,k,t is its binary mask, and Ω c i,k is the selected token set: C =reasoning, rating, format, Ω c i,k = t χ c i,k,t = 1 , c∈C. (10) Each reward is normalized independently within the K sam- ples of image i: A c i,k = R c i,k − μ c i σ c i + ε A .(11) Here, μ c i and σ c i denote the corresponding group mean and sample standard deviation. The masked advantage for each supervision type is e A c i,k,t = χ c i,k,t A c i,k .(12) Thus, reasoning and rating rewards update only their cor- responding fields, whereas the format reward applies to the complete output. 3.4 Masked Credit with GRPO We optimize the Actor with Group Relative Policy Opti- mization (GRPO) (Shao et al. 2024). LetI c contain the valid sampled outputs for supervision type c. Using the masked advantage in Eq. (12), the masked GRPO loss is L GRPO (θ) =− X c∈C 1 |I c | X (i,k)∈I c 1 |Ω c i,k | X t∈Ω c i,k ω i,k,t × min ρ i,k,t e A c i,k,t ,ρ clip i,k,t e A c i,k,t . (13) Here, ρ i,k,t is the likelihood ratio between the current Ac- tor π θ and the rollout policy π old . ρ clip i,k,t clips this ratio to [1−ε low , 1 +ε high ]. ω i,k,t corrects the token-level mismatch between the rollout distribution and π old and is capped by ω max . 4 Experiments 4.1 Experimental Settings Datasets. We train all controlled models on the training subset of the KonIQ-10k split (Hosu et al. 2020), which con- tains 7,046 in-the-wild images at 512×384 resolution. We evaluate on the KonIQ-10k test split (2,010 images) as the in-distribution authentic benchmark. For out-of-distribution (OOD) evaluation, we use the authentic distortion datasets SPAQ (11,125 images) (Fang et al. 2020) and LIVE In the Wild (LIVE-W; 1,162 images) (Ghadiyaram and Bovik 2015), as well as the AI-generated image quality dataset AGIQA-3K (2,982 images) (Li et al. 2024). We further in- clude the synthetic distortion datasets KADID-10k (10,125 images) (Lin, Hosu, and Saupe 2019) and CSIQ (866 im- ages) (Larson and Chandler 2010). Model Backbones. The Actor is initialized from Qwen3.5-4B (Qwen Team 2026), selected for its balance of visual reasoning and efficiency. The Editor uses FLUX.2 [klein] 4B (Black Forest Labs 2025) with a four-step LoRA for fast and efficient editing. It runs in BF16 with classifier- free guidance 1.0 and sigma schedule [1.0, 0.75, 0.5, 0.25]. The Judge uses a Qwen3.5-4B checkpoint trained for 5 epochs on KonIQ-7k with rating supervision, selected for the same balance of visual reasoning and efficiency. On a single NVIDIA RTX A6000, the Editor’s mean inference time is 0.902 s per image, while Judge inference takes 0.800 s per im- age. During GRPO optimization, only the Actor is updated; the frozen Editor and Judge provide offline supervision and receive no gradients. Implementation Details. We train Qwen3.5-4B (Qwen Team 2026) for five epochs on eight 48-GB NVIDIA RTX A6000 GPUs. Each rank processes N = 6 images and sam- ples K = 6 completions per image. The maximum comple- tion length is 192 tokens. We use AdamW (Loshchilov and Hutter 2019) with (β 1 ,β 2 ) = (0.9, 0.95), ε Adam = 10 −8 , a learning rate of 10 −6 , weight decay of 0.1, cosine scheduling without warmup, and gradient-norm clipping at 1.0. We use symmetric policy clipping with (ε low ,ε high ) = (0.20, 0.20), no hard clipping of advantages, and token-level importance- ratio truncation at ω max = 2. Rollouts use temperature 0.7, AuthenticAI-generatedSyntheticAverage KonIQSPAQLIVE-W AGIQA-3K KADID-10kCSIQAvg. MethodPLCC SRCC PLCC SRCC PLCC SRCC PLCC SRCC PLCC SRCC PLCC SRCC PLCC SRCC Hand-crafted NIQE (2012)0.533 0.530 0.679 0.664 0.493 0.449 0.560 0.533 0.468 0.405 0.718 0.628 0.575 0.535 BRISQUE (2012) 0.225 0.226 0.490 0.406 0.361 0.313 0.541 0.497 0.429 0.356 0.740 0.556 0.464 0.392 Deep-learning-based NIMA (2018)0.896 0.859 0.838 0.856 0.814 0.771 0.715 0.654 0.532 0.535 0.695 0.649 0.748 0.721 DBCNN (2020)0.884 0.875 0.812 0.806 0.773 0.730 0.641 0.648 0.497 0.484 0.586 0.572 0.699 0.686 MUSIQ (2021)0.924 0.929 0.868 0.863 0.789 0.830 0.722 0.630 0.575 0.556 0.771 0.710 0.775 0.753 MANIQA (2022) 0.849 0.834 0.768 0.758 0.849 0.832 0.723 0.636 0.499 0.465 0.623 0.627 0.719 0.692 CLIP-IQA+ (2023) 0.909 0.895 0.866 0.864 0.832 0.805 0.736 0.685 0.653 0.654 0.772 0.719 0.795 0.770 MLLM-based: SFT training C2Score (2024)0.923 0.910 0.867 0.860 0.786 0.772 0.777 0.671 0.500 0.453 0.735 0.705 0.765 0.729 Q-Align (2023)0.941 0.940 0.886 0.887 0.853 0.860 0.772 0.735 0.674 0.684 0.671 0.737 0.800 0.807 DeQA (2025)0.953 0.941 0.895 0.896 0.892 0.879 0.809 0.729 0.694 0.687 0.787 0.744 0.838 0.813 MLLM-based: RL training Q-Insight (2025a) 0.918 0.895 0.903 0.903 0.870 0.839 0.816 0.766 0.702 0.702 0.685 0.640 0.816 0.791 VQ-R1 (2025)0.886 0.919 0.867 0.887 0.817 0.869 0.744 0.718 0.635 0.640 0.709 0.721 0.776 0.792 MR-IQA (2026)0.949 0.931 0.892 0.897 0.899 0.883 0.804 0.732 0.672 0.683 0.767 0.732 0.831 0.810 MLLM-based: Tool-augmented training Zoom-IQA † (2026) 0.938 0.922 0.902 0.900 0.887 0.870 0.816 0.765 0.701 0.700 0.797 0.754 0.840 0.819 MR-IQA-2 (Ours) 0.937 0.917 0.900 0.899 0.893 0.863 0.809 0.739 0.667 0.669 0.824 0.785 0.838 0.812 MR-IQA-2 ∗ (Ours) 0.925 0.904 0.901 0.900 0.881 0.847 0.826 0.768 0.665 0.676 0.840 0.801 0.840 0.816 Table 1: Rating performance comparison. Each dataset reports PLCC↑ and SRCC↑ 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. Pretrained JudgeQ-Insight as JudgeGPT-5.6 Sol as Judge (%) Actor Editor FLUX.2 [klein]Mage-FlowFLUX.2 [klein]Mage-FlowFLUX.2 [klein] Qwen3.5-4B (Baseline)−0.137−0.144−0.130−0.0804.19% Q-Insight+0.213+0.142+0.138+0.13914.41% MR-IQA-2 (Ours)+0.789+0.468+0.561+0.37681.40% Table 2: Reasoning performance validation. The first four columns report mean quality gain ∆s (higher is better; Eq. (5)); GPT-5.6 Sol uses 3-Alternative Forced Choice (3AFC). We compare the Qwen3.5-4B baseline (Qwen Team 2026), Q-Insight (Li et al. 2025a), and MR-IQA-2 (ours). The 4B four-step LoRA Editors are FLUX.2 [klein] (Black Forest Labs 2025) and Mage- Flow (Zhang et al. 2026). Judges are our pretrained Qwen3.5-4B Judge (Qwen Team 2026), Q-Insight (Li et al. 2025a), and GPT-5.6 Sol (OpenAI 2026). Evaluation uses a fixed 5,866-image subset formed by randomly selecting 1,000 images from each test set, except CSIQ, for which all 866 images are used. top-p 1.0, top-k 20, and a presence penalty of 1.5. Actor- only training takes approximately 2 hours per epoch. For the full MR-IQA-2 framework, four GPUs are allocated to Actor training and four to Editor–Judge inference, increasing the per-epoch time to approximately 12 hours. 4.2 Rating Performance Compared methods. Table 1 compares MR-IQA-2 with representative BIQA methods across six benchmarks. The hand-crafted group includes NIQE (Mittal, Soundarara- jan, and Bovik 2012) and BRISQUE (Mittal, Moorthy, and Bovik 2012). Deep-learning-based methods include NIMA (Talebi and Milanfar 2018), DBCNN (Zhang et al. 2020), MUSIQ (Ke et al. 2021), MANIQA (Yang et al. 2022), and CLIP-IQA+ (Wang, Chan, and Loy 2023). Among MLLM-based methods, SFT approaches include C2Score (Zhu et al. 2024), Q-Align (Wu et al. 2023), and DeQA (You et al. 2025), while RL approaches include Q- Insight (Li et al. 2025a), VQ-R1 (Wu et al. 2025), and MR-IQA (Li et al. 2026). We further compare with the tool-augmented RL method Zoom-IQA (Liang et al. 2026). For MR-IQA-2, we report an active-vision checkpoint and a frozen-vision variant, denoted by ∗ . Efficiency and performance. MR-IQA-2 is designed pri- marily to improve reasoning faithfulness rather than maxi- mize rating. Nevertheless, it achieves competitive rating per- formance. (1) Zoom-IQA (Liang et al. 2026) uses a two- stage SFT–RL pipeline with additional annotations, whereas MR-IQA-2 uses only RL training without additional data. The frozen-vision variant reaches an average PLCC/SRCC of 0.840/0.816, comparable to Zoom-IQA’s 0.840/0.819. (2) Freezing the vision encoder and aligner is particularly ef- fective on AI-generated and synthetic datasets. The frozen variant achieves the best CSIQ PLCC/SRCC of 0.840/0.801, whereas active visual training yields its main gains on the authentic datasets. 4.3 Faithful Reasoning Performance Quality Gain as a Proxy for Reasoning Faithfulness. Prior work, including Q-Insight (Li et al. 2025a) and VisualQuality-R1 (Wu et al. 2025), typically evaluates rea- soning through human inspection of a small number of ex- amples or indirectly through improvements in rating perfor- mance. These practices lack a unified evaluation criterion. In this work, we instead use the image-quality gain produced by reasoning-guided editing as the evaluation criterion (Eq. (6)), rather than rating performance. We assume that faithful rea- soning should identify quality-limiting factors whose correc- tion improves the given image. Cross-editor/judge stability. Table 2 examines whether the Actor overfits to the frozen Editor–Judge during train- ing. We first replace the FLUX.2 [klein] Editor (Black Forest Labs 2025) with Mage-Flow (Zhang et al. 2026). MR-IQA-2 retains larger quality gains than the competing Actors, in- dicating that its reasoning is not specific to a single Editor. We then replace the pretrained Judge with Q-Insight (Li et al. 2025a) and GPT-5.6 Sol (OpenAI 2026). These alter- native Judges preserve the direction of the measured gains. These results indicate that the Actor learns generalizable quality knowledge rather than exploiting idiosyncrasies of the training-time Editor or Judge. Quality and behavior analysis. Additional analyses (in Appendix Figure 4) provide four observations. (a) Across all datasets, lower-quality images obtain larger quality gains. (b) Changes in sharpness are strongly associated with Judge preference. (c) Higher-quality images tend to remain closer to their originals after editing. (d) Training shifts the Actor toward quality-relevant reasoning tokens. 4.4 Ablation Study Evaluation. Table 3 reports ablations on rating align- ment and reasoning-guided quality improvement. Rating- only training already achieves competitive rating alignment, with an average PLCC/SRCC of 0.836/0.815. However, its mean quality gain is −0.260, below the baseline result of −0.136. Rating performance alone therefore does not guar- antee faithful reasoning. At E5, the variants without and with the credit mask reach average PLCC/SRCC values of 0.834/0.812 and 0.838/0.812, respectively. Both produce positive quality gains. The variant without the mask reaches a larger raw gain of +1.442, but collapses to one normalized solution. The credit-mask variant retains 23,457 normalized solutions while achieving a gain of +1.079. Therefore, raw gain alone can favor a universal edit and does not establish image-conditioned reasoning. The final E5 comparison is di- agnostic because the two variants also differ in KL scope; controlled E1 results are reported in Appendix E. The credit mask is more beneficial early in training. At step 30 on the 200-image validation set, it reaches a PLCC/S- RCC of 0.745/0.772, compared with 0.667/0.617 without the credit mask and 0.616/0.610 for the baseline. Its qual- ity gain is lower at this stage, at 0.255 versus 0.387 without the credit mask. This suggests that the credit mask acceler- ates rating convergence but may temporarily limit reasoning improvement. Once rating alignment stabilizes, its benefit becomes smaller. This behavior warrants further study of the imbalance between rating and reasoning rewards. 4.5 Case Study Figure 2 compares reasoning-guided edits before and after training. After training, all three edits receive higher Judge ratings, and the Actor proposes more specific interventions. In the dog example, it recommends cropping the distracting foot near the image boundary to improve visual appeal. The examples also expose a semantic-preservation problem: an edited image may no longer retain the meaning of the origi- nal. In the first row, the Editor changes a night sky into a blue daytime sky. The boundary between quality enhancement and semantic alteration therefore requires further study. 5 Discussion Good Evidence, Bad Solution? In this work, we jointly treat evidence and solution as reasoning. However, as with human perception, a model may correctly identify why an image has poor quality yet fail to propose an effective correction. Our training necessarily encourages the model to discover effective solutions through exploration, raising the question of whether this process also improves its evidence. Maybe the evidence capability does not improve. But we argue that a solution that improves image quality through relevant visual changes indicates an improved understanding of the underlying quality factors. Explicitly modeling the causal relation between evidence and solution may provide a more reliable direction for future work. 6 Limitations and Future Work This work has two main limitations. First, the current frame- work does not yet realize fully closed-loop visual reflection. The edited image is evaluated by the Judge but is not re- turned to the Actor for further visual reasoning. Second, using separate Actor, Editor, and Judge models introduces computational and parameter redundancy. Future work will close the reflection loop by feeding the edited image back to the Actor, enabling direct comparison between the original Rating alignment AuthenticAI-generatedSyntheticAverage KonIQSPAQLIVE-W AGIQA-3K KADID-10kCSIQAvg. MethodPLCC SRCC PLCC SRCC PLCC SRCC PLCC SRCC PLCC SRCC PLCC SRCC PLCC SRCC Baseline0.553 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-only0.952 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 Credit mask (E5)0.937 0.917 0.900 0.899 0.893 0.863 0.809 0.739 0.667 0.669 0.824 0.785 0.838 0.812 Subset quality gains MethodOriginal meanEdited meanMean gainPositive gain (%) Baseline2.8442.708−0.13640.664% Rating-only2.8442.584−0.26030.686% Without credit mask (E5)2.8444.286+1.44299.997% Credit mask (E5)2.8443.923+1.07998.813% 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. Input ImageBefore TrainingAfter Training Evidence: Significant noise and grain due to high ISO in low-light conditions. Solution: Apply a denoising filter to reduce the visible grain while preserving edge details, improve the visibility of textures on the dog's fur and the wooden planks. Evidence: Low resolution, motion blur poor lighting around the dog's face and background details Solution: Brighter light would enhance quality. Cropping to remove peripheral elements like the foot would improve composition. MOS:3.38 Judge rating:3.86 Judge rating:3.17 Judge rating:4.97 Evidence: Significant noise and grain, particularly in the dark background. Solution: Apply a denoising filter to reduce grain while preserving edge details, followed by a slight sharpening pass to enhance texture. Evidence: Blurred petals and low contrast in the central flower region. Solution: Taking a clearer shot in brighter light would enhance quality. Adjusting focus could sharpen the flowers. MOS:3.63 Judge rating:4.15 Judge rating:3.15Judge rating:5.00 Evidence: Severe underexposure, resulting in a very dark scene,the foreground and background are lost in shadow, obscuring details. Solution: Increase exposure to brighten the dark areas and recover details while preserving the tower. Evidence: Low visibility and lack of detail,around the central illuminated structure Solution: Taking a clearer shot in brighter conditions would enhance quality. Stabilizing the focusing on the lit structures could yield better. MOS:1.77 Judge rating:2.79Judge rating:2.84 Judge rating:4.86 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. and edited images. We will also explore smaller models and parameter sharing across modules to improve efficiency. 7 Conclusion In this work, we aim to supervise the faithfulness of image- quality reasoning, moving BIQA beyond plausible explana- tions toward visually verifiable understanding. We use quality gain as an operational measure of reasoning faithfulness and propose MR-IQA-2, an Actor–Editor–Judge framework. The Actor’s reasoning guides the Editor, while the Judge evalu- ates the resulting quality change to test whether the identified factors are supported. Fine-grained credit assignment further decouples reasoning and rating supervision to reduce reward ambiguity. MR-IQA-2 retains competitive rating alignment while producing more reliable and actionable quality reason- ing. More broadly, our framework provides a new perspec- tive on evaluating reasoning faithfulness, links blind assess- ment to reference-based verification through intervention- generated image pairs, and suggests a path toward modeling human visual preferences and aesthetic judgments. References Agnolucci, L.; Galteri, L.; Bertini, M.; and Del Bimbo, A. 2024. ARNIQA: Learning Distortion Manifold for Image Quality Assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 189–198. Bai, S.; Cai, Y.; Chen, R.; Chen, K.; Chen, X.; Cheng, Z.; Deng, L.; Ding, W.; Gao, C.; Ge, C.; et al. 2025a. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631. Bai, S.; Chen, K.; Liu, X.; Wang, J.; Ge, W.; Song, S.; Dang, K.; Wang, P.; Wang, S.; Tang, J.; et al. 2025b. Qwen2.5-VL Technical Report. arXiv preprint arXiv:2502.13923. Black Forest Labs. 2025. FLUX.2: Frontier Visual Intelli- gence. https://bfl.ai/blog/flux-2. Chen, C.; Mo, J.; Hou, J.; Wu, H.; Liao, L.; Sun, W.; Yan, Q.; and Lin, W. 2024. TOPIQ: A Top-Down Approach from Se- mantics to Distortions for Image Quality Assessment. IEEE Transactions on Image Processing, 33: 2404–2418. Fang, Y.; Zhu, H.; Zeng, Y.; Ma, K.; and Wang, Z. 2020. Perceptual quality assessment of smartphone photography. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 3677–3686. Ghadiyaram, D.; and Bovik, A. C. 2015. Live in the wild image quality challenge database. Online: http://live. ece. utexas. edu/research/ChallengeDB/index. html [Mar, 2017]. Hosu, V.; Lin, H.; Sziranyi, T.; and Saupe, D. 2020. KonIQ- 10k: An Ecologically Valid Database for Deep Learning of Blind Image Quality Assessment. IEEE Transactions on Image Processing, 29: 4041–4056. Ke, J.; Wang, Q.; Wang, Y.; Milanfar, P.; and Yang, F. 2021. MUSIQ: Multi-scale Image Quality Transformer. In Pro- ceedings of the IEEE/CVF International Conference on Com- puter Vision, 5148–5157. Larson, E. C.; and Chandler, D. M. 2010. Most apparent dis- tortion: full-reference image quality assessment and the role of strategy. Journal of electronic imaging, 19(1): 011006– 011006. Li, C.; Zhang, Z.; Wu, H.; Sun, W.; Min, X.; Liu, X.; Zhai, G.; and Lin, W. 2024. AGIQA-3K: An Open Database for AI- Generated Image Quality Assessment. IEEE Transactions on Circuits and Systems for Video Technology, 34(8): 6833– 6846. Li, W.; Zhang, X.; Zhao, S.; Zhang, Y.; Li, J.; Zhang, L.; and Zhang, J. 2025a. Q-Insight: Understanding Image Quality via Visual Reinforcement Learning. arXiv preprint arXiv:2503.22679. Li, Y.; Lin, Y.; Sun, Z.; Yang, Y.-H.; Miyoshi, K.; Chu, C.; and Nishida, S. 2026. MR-IQA: A Unified Margin View of Regression and Ranking for Blind Image Quality Assess- ment. arXiv:2606.29760. Li, Y.; Sun, Z.; Chen, Y.-j.; and Nishida, S. 2025b. Building Reasonable Inference for Vision-Language Models in Blind Image Quality Assessment. In International Conference on Neural Information Processing, 283–295. Springer. Li, Y.; Yu, Y.; Lin, Y.; Yang, Y.-H.; Chu, C.; and Nishida, S. 2025c. Guiding Perception-Reasoning Closer to Hu- man in Blind Image Quality Assessment. arXiv preprint arXiv:2512.16484. Liang, G.; Wang, J.; Wu, Z.; Zhou, S.; and Loy, C. C. 2026. Zoom-IQA: Image Quality Assessment with Reliable Region-Aware Reasoning. arXiv preprint arXiv:2601.02918. Lin, H.; Hosu, V.; and Saupe, D. 2019. KADID-10k: A large- scale artificially distorted IQA database. In 2019 Eleventh International Conference on Quality of Multimedia Experi- ence (QoMEX), 1–3. IEEE. Loshchilov, I.; and Hutter, F. 2019. Decoupled Weight Decay Regularization. In International Conference on Learning Representations. Mittal, A.; Moorthy, A. K.; and Bovik, A. C. 2012. No- reference image quality assessment in the spatial domain. IEEE Transactions on image processing, 21(12): 4695–4708. Mittal, A.; Soundararajan, R.; and Bovik, A. C. 2012. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3): 209–212. OpenAI. 2026. GPT-5.6 Sol Model. https://developers. openai.com/api/docs/models/gpt-5.6-sol. Qin, G.; Zhang, J.; He, C.; Fu, Y.; Liang, J.; Wu, T.; and Zhang, L. 2026. Tool-IQA: Augmenting Image Quality Assessment with Simple Tools. arXiv preprint arXiv:2606.16082. Qwen Team. 2026. Qwen3.5: Towards Native Multimodal Agents. Shao, Z.; Wang, P.; Zhu, Q.; Xu, R.; Song, J.; Bi, X.; Zhang, H.; Zhang, M.; Li, Y.; Wu, Y.; et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open lan- guage models. arXiv preprint arXiv:2402.03300. Székely, G. J.; Rizzo, M. L.; and Bakirov, N. K. 2007. Mea- suring and Testing Dependence by Correlation of Distances. The Annals of Statistics, 35(6): 2769–2794. Talebi, H.; and Milanfar, P. 2018. NIMA: Neural image assessment. IEEE transactions on image processing, 27(8): 3998–4011. Wang, J.; Chan, K. C.; and Loy, C. C. 2023. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI conference on artificial intelligence, volume 37, 2555–2563. Wang, Z.; Simoncelli, E. P.; and Bovik, A. C. 2003. Multi- Scale Structural Similarity for Image Quality Assessment. In Proceedings of the 37th Asilomar Conference on Signals, Systems and Computers, volume 2, 1398–1402. Wu, H.; Zhang, Z.; Zhang, E.; Chen, C.; Liao, L.; Wang, A.; Xu, K.; Li, C.; Hou, J.; Zhai, G.; et al. 2024. Q-Instruct: Improving Low-Level Visual Abilities for Multi-Modality Foundation Models. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, 25490– 25500. Wu, H.; Zhang, Z.; Zhang, W.; Chen, C.; Liao, L.; Li, C.; Gao, Y.; Wang, A.; Zhang, E.; Sun, W.; et al. 2023. Q-align: Teaching lmms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090. Wu, T.; Zou, J.; Liang, J.; Zhang, L.; and Ma, K. 2025. VisualQuality-R1: Reasoning-Induced Image Quality As- sessment via Reinforcement Learning to Rank. Advances in Neural Information Processing Systems, 38: 88167–88190. Yang, S.; Wu, T.; Shi, S.; Lao, S.; Gong, Y.; Cao, M.; Wang, J.; and Yang, Y. 2022. Maniqa: Multi-dimension attention network for no-reference image quality assessment. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 1191–1200. You, Z.; Cai, X.; Gu, J.; Xue, T.; and Dong, C. 2025. Teach- ing large language models to regress accurate image quality scores using score distribution. In Proceedings of the Com- puter Vision and Pattern Recognition Conference, 14483– 14494. You, Z.; Li, Z.; Gu, J.; Yin, Z.; Xue, T.; and Dong, C. 2024. Depicting beyond scores: Advancing image quality assess- ment through multi-modal language models. In European Conference on Computer Vision, 259–276. Springer. Yu, Q.; Zhang, Z.; Zhu, R.; Yuan, Y.; Zuo, X.; et al. 2025. DAPO: An Open-Source LLM Reinforcement Learning Sys- tem at Scale. In Advances in Neural Information Processing Systems, volume 38. Zhang, R.; Isola, P.; Efros, A. A.; Shechtman, E.; and Wang, O. 2018. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 586–595. Zhang, W.; Ma, K.; Yan, J.; Deng, D.; and Wang, Z. 2020. Blind Image Quality Assessment Using A Deep Bilinear Convolutional Neural Network. IEEE Transactions on Cir- cuits and Systems for Video Technology, 30(1): 36–47. Zhang, W.; Zhai, G.; Wei, Y.; Yang, X.; and Ma, K. 2023. Blind image quality assessment via vision-language corre- spondence: A multitask learning perspective. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 14071–14081. Zhang, X.; Zhang, P.; Zheng, S.; Guo, J.; Jia, Z.; Shen, Y.; Guo, X.; Luo, Y.; Li, J.; Xie, W.; Pu, F.; Zhang, X.; Zhang, K.; Guo, Z.; Bi, T.; Gui, D.; Liu, Z.; Wen, Z.; Zheng, Z.; Yang, S.; Li, X.; Wang, J.; Li, B.; and Lu, Y. 2026. Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing. arXiv preprint arXiv:2607.19064. Zhu, H.; Wu, H.; Li, Y.; Zhang, Z.; Chen, B.; Zhu, L.; Fang, Y.; Zhai, G.; Lin, W.; and Wang, S. 2024. Adaptive image quality assessment via teaching large multimodal model to compare. Advances in Neural Information Processing Sys- tems, 37: 32611–32629. In this appendix, we provide technical details and more comprehensive analyses of our proposed method. It com- prises Section A, Backbone, Algorithm and Hyperparame- ters; Section B, Visual Quality Feature Analysis; Section C, Discussion on the Editor Proxy; Section D, Joint Rating, Reasoning, and Diversity Analysis; and Section E, Detailed Credit-Mask Audit. A Backbone, Algorithm and Hyperparameters A.1 Dataset Protocol. Our training and evaluation settings follow the same con- figuration as the main paper. We train on the KonIQ-10k training subset (Hosu et al. 2020) and use a fixed set of 200 images randomly sampled from the KonIQ-10k test split to monitor early-stage convergence. Final testing is con- ducted on the KonIQ-10k test split, SPAQ (Fang et al. 2020), LIVE-W (Ghadiyaram and Bovik 2015), AGIQA-3K (Li et al. 2024), KADID-10k (Lin, Hosu, and Saupe 2019), and CSIQ (Larson and Chandler 2010). At the time of manuscript completion, we did not have access to Zoom-IQA (Liang et al. 2026) and therefore could not reproduce it under our protocol. Tool-IQA (Qin et al. 2026) follows a different training–testing configuration; con- sequently, we cannot report directly comparable Tool-IQA results for some experiments. For the cross Actor–Judge–Editor evaluation reported in Table 2 of the main paper, we use a fixed 5,866-image subset, randomly sampling 1,000 images from each test set except CSIQ, for which all 866 images are used. This subsampling is necessary for computational efficiency: evaluating the full test sets would involve approximately 23,000 images, each requiring four editing operations and three Judge evaluations, resulting in a prohibitively heavy inference cost. A.2 Candidate Backbones. The Qwen series has been widely adopted in recent BIQA studies due to its open-source availability, strong multimodal capability, and computational efficiency. For example, exist- ing methods such as Q-Insight (Li et al. 2025a), VQ-R1 (Wu et al. 2025), and Zoom-IQA (Liang et al. 2026) employ Qwen2.5-VL-7B (Bai et al. 2025b) as their backbone model. However, despite its strong performance, the 7B-scale model still introduces considerable computational overhead, which limits its practical deployment and training efficiency. Therefore, in this work, we investigate more lightweight backbone configurations with 2B and 4B parameters. Con- sidering future scalability and the continuous evolution of multimodal large language models, we further evaluate the latest Qwen3-VL and Qwen3.5-VL architectures. We con- duct comprehensive backbone comparisons among 2B/4B variants of Qwen3 and Qwen3.5 to identify an optimal trade- off between training efficiency and performance. Qwen3-VL versus Qwen3.5. The T1–T2 comparison in Table 4 shows that, under matched frozen-vision DAPO set- tings, Qwen3.5-2B outperforms Qwen3-VL-2B on five of six datasets for both PLCC and SRCC, and on all six in MAE. It also requires only 0.81× the training time per epoch. 2B versus 4B. The T6–T7 comparison in Table 4 shows that the 4B model converges faster in rating performance. By Epoch 2, T7 reaches a validation PLCC/SRCC/MAE of 0.929/0.919/0.566, compared with 0.915/0.904/0.744 for the 2B T6; by Epoch 3, T7 further improves to 0.935/0.924/0.228. The 4B model also produces more di- verse outputs, with 83 distinct ratings, 200 unique comple- tions, and a 41.5% U-score, compared with a 39.0% U-score for T6. We further observe that its response lengths are bet- ter aligned with our objective of eliciting detailed quality reasoning. A.3 Optimization Algorithms: GRPO or DAPO? Compared with GRPO (Shao et al. 2024), DAPO (Yu et al. 2025) dynamically resamples low-variance groups, thereby increasing within-group reward diversity and preventing pol- icy advantages from vanishing prematurely. We therefore adopted DAPO in most of our early experiments. Across three epochs, 3,282 of 20,832 GRPO learner groups (15.75%) have zero within-group reward variance and therefore zero policy advantage. DAPO resampling re- duces the proportion of zero-advantage learner groups to 0%. Comparing T2 and T4 in Table 4, DAPO improves the six-dataset average PLCC/SRCC by 0.006/0.008 and reduces MAE by 0.138 at Epoch 3, with improvements on five, six, and six datasets, respectively. Despite its better performance, DAPO requires 1.866× more sampled trajectories and increases the training time per epoch from 0.81 to 1.13 hours (approximately 39.5%). This additional training-time cost is not affordable in our training environment, especially when scaling the model to 4B parameters or beyond. Therefore, we use GRPO as the default optimization algorithm in the final framework. A.4 Hyper-Parameter Settings. Group Size: 6 versus 48. MR-IQA (Li et al. 2026) reports its best performance with local groups of six, where rewards are computed by comparing samples only within each six-image group. Table 4 compares this local- 6 setting (T6) with a larger group of 48 (T5). The large group achieves better in-domain validation PLCC/S- RCC/MAE (0.929/0.921/0.261 versus 0.917/0.906/0.587), whereas local-6 achieves higher six-dataset average PLCC/S- RCC (0.835/0.813 versus 0.828/0.809). Because the local-6 setting provides stronger cross-dataset correlation, we adopt a group size of six. Vision Modules: Frozen versus Active. The T2–T5 com- parison in Table 4 shows that active visual training per- forms better on authentic datasets, achieving PLCC/S- RCC values of 0.952/0.938 on KonIQ and 0.897/0.877 on LIVE-W, compared with 0.925/0.904 and 0.881/0.847 for the frozen variant. In contrast, freezing the vision en- coder and aligner performs better on AGIQA-3K and the synthetic datasets: it achieves 0.826/0.768 on AGIQA-3K, 0.665/0.676 on KADID-10k, and 0.840/0.801 on CSIQ, com- pared with 0.816/0.747, 0.661/0.668, and 0.771/0.746 under active visual training. Consequently, the frozen variant yields VariantIter. Time (h/ep.)Val. P/S/MU-score (%)Gen. P/S/M (a) Qwen3-VL-2B T1: DAPO i1, frozen, global11.38 0.831/0.817/0.86812.00.809/0.787/0.936 (b) Qwen3.5-2B T2: DAPO i1, frozen, global11.13 0.890/0.881/0.69136.00.835/0.808/0.740 T3: DAPO i4, frozen, global42.47 0.910/0.904/0.26521.50.834/0.809/0.541 T4: GRPO i1, frozen10.81 0.886/0.883/0.84534.00.829/0.800/0.878 T5: DAPO i1, visual, global11.30 0.929/0.921/0.26135.00.828/0.809/0.513 T6: DAPO i1, visual, local-six 11.54 0.917/0.906/0.58739.00.835/0.813/0.684 (c) Qwen3.5-4B, local-six T7: DAPO i1, visual, local-six 12.56 0.935/0.924/0.22841.5– Table 4: Actor-only ablations at their reported endpoints. P/S/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. 0.76 0.80 0.84 0.88 0.92 (a) PLCC ↑ 0.75 0.80 0.85 0.90 (b) SRCC ↑ 123 Epoch 0.2 0.4 0.6 0.8 1.0 (c) MAE ↓ 123 Epoch 0 10 20 30 40 (d) U-score (%) T1: 3VL-2B DAPO T5: 2B visual-global T2: 2B DAPO i1 T6: 2B visual-local T3: 2B DAPO i4 T7: 4B local-six T4: 2B GRPO i1 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. higher six-dataset average PLCC/SRCC (0.840/0.816 versus 0.833/0.812), indicating stronger overall generalization. Rating Diversity. Zoom-IQA (Liang et al. 2026) reports score-space collapse when using a ranking reward, with only a 2.04% unique-score ratio. Across T1–T7 in Table 4, the E3 U-scores remain between 12.0% and 41.5%, corresponding to 24–83 distinct ratings on the 200-image validation set. Thus, we do not observe rating-diversity collapse under our current settings. Repeated Sampling Updates. MR-IQA (Li et al. 2026) reuses each rollout sample for four policy updates. We eval- uate whether these repeated updates remain effective by comparing T2 and T3 in Table 4. Four updates improve validation PLCC/SRCC/MAE from 0.890/0.881/0.691 to 0.910/0.904/0.265. However, the six-dataset average PLC- C/SRCC remain nearly unchanged (0.835/0.808 versus 0.834/0.809), while training time increases from 1.13 to 2.47 hours per epoch. Therefore, one policy update per rollout is sufficient under our current settings. B Visual Quality Feature Analysis In our framework, the FLUX.2 [klein] Editor (Black For- est Labs 2025) produces an edited image conditioned on the quality reasoning for each input. The resulting original– edited pair bridges single-image quality assessment with reference-based quality assessment, enabling us to analyze how the visual intervention relates to the predicted qual- Figure 4: Quality gain versus visual change scale. The panels relate the Judge-observed quality gain ∆s to PSNR, MS- SSIM (Wang, Simoncelli, and Bovik 2003), and LPIPS (Zhang et al. 2018) between the original and edited images. Colors denote datasets, curves show smoothed trends, and ρ denotes Spearman correlation. ity improvement. In this section, we investigate three ques- tions: (1) whether larger visual changes lead to greater quality gains; (2) how the distributions of low-level visual features change after editing; and (3) what behavioral preferences emerge from the overall framework. B.1 Visual Change and Quality Gains Visual change and quality improvement. Figure 4 re- lates the Judge-observed quality gain to three reference- based measures of the visual change between the original and edited images. Quality gain is generally negatively cor- related with PSNR (ρ = −0.33 to +0.01) and MS-SSIM (ρ =−0.49 to−0.28), and positively correlated with LPIPS (ρ = 0.37 to 0.56). Because larger changes correspond to lower PSNR/MS-SSIM and higher LPIPS, these consis- tent directions indicate that larger editing changes are gen- erally associated with greater image-quality improvements. SPAQ deviates from the other datasets across multiple mea- sures. Its quality-gain correlation is nearly zero for PSNR (ρ = +0.01) and is the weakest in magnitude for both MS- SSIM (ρ = −0.28) and LPIPS (ρ = +0.37). A similar pattern appears in rating performance: while the rating cor- relations on the other datasets change during training, the SPAQ correlation remains close to 0.90 with only minor vari- ation. This stability suggests that SPAQ is less sensitive to the training-induced changes observed on the other benchmarks. B.2 Low-Level Visual Feature Patterns Human-like Judge feedback. Figure 5 shows that the feature-dependent curves produced by the original-image Judge broadly follow the corresponding human-MOS curves across the six datasets. Their similar shapes and directions, particularly for sharpness and contrast, indicate that the Judge captures human-like quality preferences over low-level vi- sual attributes. This agreement supports its use as a qualified source of human-like feedback for the original–edited image pairs. Cross-dataset feature patterns. Among luminance, con- trast, saturation, sharpness, colorfulness, and entropy, sharp- ness exhibits the clearest cross-dataset trend. Both human MOS and Judge scores generally increase with sharpness, whereas the other attributes show weaker, nonlinear, or dataset-dependent patterns. This result suggests that sharp- ness is a broadly shared quality cue, while no single low-level attribute fully explains perceptual quality across all datasets. Edited-feature distributions. After editing, the images re- tain broad distributions across all six low-level attributes rather than collapsing to fixed feature values. The frame- work therefore does not appear to overfit a single low-level pattern when improving image quality. Nevertheless, Fig- ure 5 examines each attribute marginally. The joint distri- bution of low-level attributes, and its relation to the human visual system and learned perceptual representations such as CLIP-IQA (Wang, Chan, and Loy 2023), warrants further investigation. B.3 Framework Behavior Preference Source quality, visual change, and quality gain. Fig- ure 6(a) shows that lower-MOS images generally receive larger Judge-rated quality gains. Panel (c) further shows that lower-MOS images tend to have lower similarity between the original and edited images, indicating larger visual changes. Together with the observation in Section B.1 that larger vi- sual changes are associated with higher quality gains, these results suggest a plausible behavior: for lower-quality inputs, the model tends to modify more visual attributes, which in turn leads to greater quality improvements. This behavior aligns with the intuition that lower-quality images provide more room for correction and supports the intended opera- 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; ρ and dCor denote Spearman and distance correlations (Székely, Rizzo, and Bakirov 2007). tion of our framework, although the observed associations do not by themselves establish causality. Preferences over feature changes. Unlike Section B.2, which examines the absolute values of low-level attributes, Figure 6(b) relates their changes after editing to the resulting quality gain. Changes in sharpness retain the strongest and most consistent relationship with quality improvement. For entropy and saturation, the authentic and synthetic datasets form different response patterns, indicating that the preferred corrections depend on the degradation domain. SPAQ again exhibits signals that are less consistent with the other datasets, in agreement with the dataset-specific behavior observed in Section B.1. Behavioral adaptation and concentration. Figure 6(d) reflects how the model adjusts its reasoning vocabulary in response to Judge feedback. After training, terms such as natural and realistic become prominent, revealing a learned preference for perceptually plausible corrections. At the same time, restore alone accounts for 20.6% of the displayed token frequency, and the top-10 token share rises to 44.1%. This concentration suggests an emerging tendency to overfit a small set of solution patterns, despite the broader behavioral alignment induced by the framework. C Discussion on the Editor Proxy This work does not disentangle the Editor’s instruction- following capability from semantic preservation, creating (c). MOS vs. Reference Metrics (b). Quality Gains vs. Low-level feature (d). Token Frequency Change (a). Quality Gains vs. MOS 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. two attribution risks. First, quality may degrade because of Editor limitations rather than an incorrect Actor instruction. Second, the edited image may become semantically incon- sistent with the input, weakening quality change as a proxy for reasoning faithfulness. Future work should use indepen- dent human or human-like evaluation to assess instruction following, semantic consistency, and perceptual quality sep- arately, yielding more reliable feedback and a more faithful framework. D Joint Rating, Reasoning, and Diversity Analysis Evaluation protocol. We jointly examine rating general- ization, functional reasoning, and language diversity. Rating results are six-dataset averages over 28,270 test images. Func- tional reasoning is measured by the Judge-score gain ∆s on the same 5,843 image keys for every Actor, using a fixed FLUX Editor and E5 Judge. This common-case evaluation excludes 23 Mask generations that reached the length limit. Language diversity is evaluated on 23,599 common success- ful inputs. Evidence and solution uniqueness are exact-string statistics, with the top-1 share measuring the concentration of the most frequent output. Results should therefore be com- pared within each metric and its stated sample scope. Dataset-level rating performance. Table 5 provides the complete cross-dataset results. In the controlled E1 compar- ison, Without Mask performs best on KonIQ, SPAQ, and AGIQA-3K and has the lowest average MAE. Mask, no KL has lower average correlation. Adding KL to Mask recovers most of this gap and gives the strongest LIVE-W result. At E5, Without Mask is stronger on KADID-10k and slightly higher in CSIQ SRCC, whereas Mask has the best average PLCC and is stronger on most authentic and AI-generated bench- marks. Despite their similar average correlation, Without Mask has a substantially higher MAE (1.054 versus 0.543), exposing a calibration failure that PLCC and SRCC alone do not capture. Rating and functional reasoning. Table 6 extends the main ablation in Table 3 with output-level diagnostics. The rating-only Actor attains strong rating alignment but pro- duces edits that reduce quality on average. Among the con- trolled E1 variants, Without Mask has the strongest rating correlation, whereas Mask gives the largest quality gain. Mask + KL recovers most of the rating gap while retaining a positive mean gain. These results separate rating accuracy from the functional usefulness of reasoning. Evidence and solution diversity. Without Mask produces only three distinct evidence strings, with one string account- ing for 99.992% of the matched outputs. Its evidence has therefore collapsed even though its solutions remain more varied. Mask increases the number of unique evidence strings to 291 and yields the highest solution uniqueness. Adding KL further raises evidence uniqueness to 5,668 and lowers its top-1 share to 13.191%, at the cost of some solution diversity and edit gain. Evaluating evidence and solution separately is thus necessary: rating correlation, full-output validity, or solution diversity alone can conceal a collapsed reasoning field. Exact uniqueness remains a lexical proxy and does not by itself establish semantic diversity or faithfulness. At E5, Without Mask retains diverse evidence (74.719% unique; 0.556% top-1), while its solution uniqueness falls to 0.004% and its solution top-1 share reaches 100%. This field-wise mismatch shows why evidence and solution must be audited separately. Final E5 balance. The E5 rows in Table 6 reinforce this multi-objective interpretation. Without Mask E5 reaches a raw gain of +1.431 but maps all 28,270 inputs to one normal- ized solution in the same template family despite retaining diverse evidence. Mask E5 obtains an average PLCC/SRCC of 0.838/0.812, a mean quality gain of +1.071, and 19,618 unique evidence strings on the matched audit. It also pro- duces 23,457 normalized solutions among 28,044 successful full-audit cases, with no detected house-template outputs. The higher Without Mask reward therefore reflects a univer- sal edit rather than image-conditioned solution reasoning. Because the final runs change both credit routing and KL topology and use one seed per setting, this comparison is di- agnostic rather than a pure causal estimate of the credit mask. Section E provides the corresponding training and validation dynamics. E Detailed Credit-Mask Audit E.1 Audit Scope and Protocol We audit two complete five-epoch runs with the same Actor, data, prompt, and GRPO settings. Mask routes each reward to its supervised output with output-specific KL regulariza- tion; Without Mask broadcasts rewards over the complete output with one global KL term. Both use a KL coefficient of 0.02. The diagnostic reasoning reward bounds the raw Judge-score change: R diag i,k = sgn(∆s i,k ) 1− e −∆s 2 i,k /2 .(14) The Editor consumes only the solution without an explicit se- mantic guardrail. Both runs use six samples per image, 160- token completions, and frozen vision encoder and aligner. Each contains 1,455 steps with 144 trajectories per step (209,520 per run; 419,040 total). No step or record is missing, duplicated, invalid, or non-finite. At Without Mask steps 711 and 713, all samples are Actor-ineligible, so absent Judge changes reflect eligibility rather than numerical or service failure. Because credit and KL routing change jointly and each setting has one seed, this audit is descriptive rather than causal. E.2 Training Dynamics Figure 7 reports exact four-rank means at the retained check- points. Both configurations improve rating reward. Without Mask produces larger reasoning reward and Judge gain, but also shorter outputs and solution collapse. Mask changes more gradually: its reasoning reward rises from 0.207 to 0.307 and Judge gain from 0.586 to 0.840. Rating rewards converge by E5. KL losses are not directly comparable be- cause their token scopes differ. For Without Mask, mean reasoning reward and Judge gain increase from 0.463 and 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. MethodKonIQSPAQLIVE-W AGIQA-3K KADID-10kCSIQAverage MAE↓ Without Mask (E1) 0.952/0.938 0.901/0.898 0.897/0.877 0.816/0.747 0.661/0.668 0.771/0.746 0.833/0.812 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 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 0.676/0.684 0.815/0.786 0.834/0.812 1.054 Mask (E5)0.937/0.917 0.900/0.899 0.893/0.863 0.809/0.739 0.667/0.669 0.824/0.785 0.838/0.812 0.543 Table 6: Joint rating, functional reasoning, and diversity analysis. P/S is the six-dataset average PLCC/SRCC. ∆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). Rating Functional reasoningLanguage diversity Method P/S ↑ Mean ∆s ↑ Pos. (%) ↑ Evidence unique (%)↑ Evidence top-1 (%)↓ Solution unique (%)↑ Solution top-1 (%)↓ Baseline0.652/0.645 −0.13640.664– Rating-only0.836/0.815 −0.26030.686– Without Mask (E1) 0.833/0.812 +0.83796.9540.01399.99250.6210.822 Mask, no KL (E1) 0.822/0.798 +0.91797.7071.23321.19694.5080.085 Mask + KL (E1)0.831/0.811 +0.78895.05424.01813.19175.2280.191 Without Mask (E5) 0.834/0.812 +1.431 99.99774.719 † 0.556 † 0.004 † 100.000 † Mask (E5)0.838/0.812 +1.07198.81383.1310.40783.644 † 0.182 † 1.189 over steps 1256–1355 to 0.478 and 1.232 over steps 1356–1455. Yet their final-window slopes are−0.0108 and −0.0246; the rating-reward slope is−0.0006. The endpoint therefore masks a plateau and slight decline. E.3 Checkpoint Validation Figure 8 evaluates the retained checkpoints on the same 200 images. PLCC and SRCC improve for both runs. With- out Mask keeps a Judge gain near 1.18, while normalized- solution uniqueness falls from 91.88% at E1 to about 0.5% at E2. The audit further shows that the semantic dwelling- template rate reaches 100% at E1. Semantic collapse there- fore precedes near-exact lexical collapse. E.4 Cross-Dataset Outcomes At E5, Without Mask obtains a larger pooled quality gain than Mask (1.431 versus 1.071), although their six-dataset PLCC/SRCC averages remain close (0.834/0.812 versus 0.838/0.812). Larger gain alone therefore does not estab- lish healthier reasoning. All 28,270 Without Mask outputs share one normalized solution; Mask retains 23,457 among 28,044 successful outputs. E.5 Collapse Milestones Table 7 distinguishes semantic, phrase-level, and near-exact collapse. Under Without Mask, the semantic dwelling family exceeds 90% at step 265, whereas one normalized sentence does not exceed 90% until step 551. Duplicate matching would detect the failure 286 steps late. Mask avoids dwelling collapse, although its super-resolution and smoothing phrase backbone remains above 90% from step 407 onward. Routing and detector≥ 50%≥ 90% Sustained Without Mask: semantic dwelling236265 948–E5 Without Mask: strict target phrase356475 – Without Mask: modal normalized solu- tion 520551 973–E5 Without Mask: canonical sentence1,1331,146 – Mask: super-resolution + smoothing214260 407–E5 Table 7: Collapse milestones in global optimizer steps. Se- mantic detectors identify shared concepts before normalized near-duplicate matching identifies one sentence. E.6 Output-Length Evolution Panel (d) of Figure 7 shows that the mean training comple- tion decreases from 89.39 to 85.26 tokens under Without Mask but rises from 100.57 to 109.00 under Mask. On val- idation outputs, the corresponding solution lengths change from 32.56 to 18.00 and from 37.85 to 50.94. The former contraction is consistent with convergence to one reusable instruction. E.7 Rating and Diversity Diagnostics Without Mask retains a KonIQ PLCC/SRCC of 0.932/0.914 despite solution collapse. Ratings still vary: 2,009 of 2,010 0.24 0.32 0.40 0.48 (a) Reasoning reward ↑ 0.88 0.90 0.92 0.94 0.96 (b) Rating reward ↑ 1235 Epoch 0.6 0.8 1.0 1.2 (c) Judge gain Δs ↑ 1235 Epoch 80 88 96 104 112 (d) Completion length MaskWithout 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. 0.86 0.88 0.90 0.92 0.94 (a) Validation PLCC ↑ 0.86 0.88 0.90 0.92 0.94 (b) Validation SRCC ↑ 1235 Checkpoint epoch 0.8 0.9 1.1 1.2 (c) Validation Judge gain ↑ 1235 Checkpoint epoch 0 25 50 75 100 (d) Normalized-solution uniqueness (%) MaskWithout Mask 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. complete JSON outputs are unique, although all 2,010 solu- tions are identical after normalization. Rating correlation and full-output uniqueness therefore miss solution collapse. Eval- uation must inspect the supervised output alongside image– solution relevance and semantic preservation. E.8 Representative Validation Outputs The Without Mask E5 Actor produces different evidence and ratings for the three examples, but all solutions normalize to one dwelling instruction. Sample 1 still receives 0.865 reward because the fixed intervention raises the Judge score from 2.37 to 4.37. Thus, the reward measures edited-image quality without establishing source-image faithfulness. Sample RatingJ 0 J 1 ∆sR 11.68 2.37 4.37 2.00 0.865 23.32 3.84 4.32 0.48 0.109 32.65 3.02 4.36 1.34 0.593 Table 8: Representative Without Mask validation outputs. Different evidence and ratings lead to the same normalized solution. E.9 Causal Scope The comparison changes credit and KL routing jointly with one seed per setting. It establishes a reproducible failure mode for Without Mask/global KL, but not the responsible change. Causal attribution requires a 2×2 grid of Without Mask versus Mask and global versus component KL, with multiple seeds.