Technical Summary: MR-IQA-2

1. Problem Statement

Image Quality Assessment (IQA) has evolved from estimating synthetic distortion intensity to understanding complex, real-world image quality through multimodal large language models (MLLMs). While recent MLLM-based approaches (e.g., Q-Insight, VisualQuality-R1) improve interpretability by generating quality reasoning, a critical gap remains: reasoning faithfulness.

Existing methods often supervise reasoning solely through alignment with human quality ratings. This creates a fundamental flaw: a model may receive a high reward for an accurate rating even if its underlying reasoning is unfaithful or coincidental. Consequently, the model might attribute low quality to plausible-sounding but incorrect factors, failing to identify the true visual limitations. Furthermore, shared rewards for both rating and reasoning obscure the source of supervision, potentially reinforcing "causal reasoning collapse" where the model learns to predict scores without genuinely understanding the visual causes.

The core challenge addressed by this work is how to verify whether generated reasoning faithfully reflects the actual factors determining image quality, rather than merely producing fluent explanations that correlate with human scores.

2. Methodology: MR-IQA-2

The authors propose MR-IQA-2, a modular Actor-Editor-Judge framework designed to operationalize "reasoning-editing-reflection." The framework decouples reasoning supervision from rating supervision and introduces a verifiable signal for reasoning faithfulness.

2.1 Framework Components

1. Actor: A trainable MLLM-based Blind IQA (BIQA) model. Given an input image, it generates:
   • Reasoning: A structured output containing "Evidence" (visual factors limiting quality) and "Solution" (a plan to edit the image to address these factors).
   • Rating: A numerical quality score.
2. Editor: A frozen image-editing diffusion model (initialized from FLUX.2). It takes the Actor's "Solution" as a prompt and applies targeted edits to the original image. Crucially, the editor preserves the image's semantic content while modifying low-level quality attributes identified by the Actor.
3. Judge: A frozen MLLM trained for rating supervision. It independently evaluates the original image and the edited image, producing scores for both.

2.2 The Training Loop (Visual Reflection)

The framework operates on a trajectory where the Actor's reasoning is tested via visual intervention:

1. The Actor generates reasoning and a rating for an image.
2. The Editor modifies the image based on the Actor's "Solution."
3. The Judge compares the original and edited images, calculating a quality gain ($\Delta s = s_{edited} - s_{original}$).
4. Faithfulness Verification: If the Actor correctly identified quality-limiting factors, the corresponding edit should result in a significant quality improvement (positive $\Delta s$). If the reasoning is flawed, the edit will not improve the image (or may degrade it).

2.3 Fine-Grained Credit Assignment

To prevent the "shared reward" problem, MR-IQA-2 employs masked token-specific updates within a Group Relative Policy Optimization (GRPO) framework.

• Decoupled Rewards:
  • Reasoning Reward ($R_{reasoning}$): Derived directly from the Judge's observed quality gain ($\Delta s$). This rewards reasoning that leads to verifiable visual improvements.
  • Rating Reward ($R_{rating}$): Derived from the alignment between the predicted rating and human Mean Opinion Scores (MOS), using relative quality margins.
  • Format Reward ($R_{format}$): Ensures valid JSON output structure.
• Masked Updates: The advantage signal for each reward type is applied only to the corresponding tokens in the output sequence (e.g., the reasoning reward updates only the "Evidence" and "Solution" tokens, while the rating reward updates only the "Rating" token). This preserves the causal relation from reasoning to rating while ensuring that a correct rating does not artificially boost unfaithful reasoning.

3. Key Contributions

1. Fine-Grained Credit Assignment: The paper introduces a mechanism to decouple reasoning and rating supervision in BIQA. By using masked token-specific updates, it assigns credit to the specific output field responsible for the performance, preventing the "lucky guess" problem where accurate ratings reinforce incorrect reasoning.
2. Verifiable Supervision via Visual Reflection: Instead of relying on human-labeled reasoning (which is scarce and subjective), the framework constructs an independent supervision signal. It treats reasoning-conditioned editing as a visual intervention; the success of this intervention (measured by the Judge's quality gain) serves as a proxy for the faithfulness of the reasoning.
3. MR-IQA-2 Framework: A modular Actor-Editor-Judge architecture that operationalizes reasoning-editing-reflection. It allows for independent instantiation of components and demonstrates that visual intervention can yield richer, more faithful visual understanding beyond mere rating alignment.

4. Experimental Results

The authors evaluated MR-IQA-2 on six benchmarks, including authentic (KonIQ-10k, SPAQ, LIVE-W), AI-generated (AGIQA-3K), and synthetic (KADID-10k, CSIQ) datasets.

• Rating Performance: MR-IQA-2 achieves competitive rating alignment with humans. On the KonIQ-10k test set, it reaches a PLCC of 0.937 and SRCC of 0.917. Across all datasets, it achieves an average PLCC/SRCC of 0.838/0.812, comparable to state-of-the-art methods like Zoom-IQA (which uses additional annotations and a two-stage pipeline) and MR-IQA.
• Reasoning Faithfulness:
  • Quality Gain: Unlike baseline models (including "Rating-only" training) which often produce edits that degrade quality (negative $\Delta s$), MR-IQA-2 consistently produces positive quality gains. The "Credit Mask" variant achieved a mean quality gain of +1.079 on the audit set.
  • Stability: The framework's reasoning remains effective even when the Editor or Judge models are swapped (e.g., using Mage-Flow instead of FLUX.2, or GPT-5.6 as the Judge), indicating the Actor learns generalizable quality knowledge rather than overfitting to specific tool idiosyncrasies.
• Ablation Studies:
  • Credit Mask Necessity: Training without the credit mask leads to "solution collapse," where the model converges to a single, universal editing instruction (e.g., "apply denoising") regardless of the input image. While this yields high raw quality gains, it represents a failure of image-conditioned reasoning. The credit mask preserves diverse, image-specific solutions while maintaining rating accuracy.
  • Rating vs. Reasoning: Models trained only on rating alignment achieve high correlation but fail to improve image quality through editing (mean gain of -0.260), confirming that rating accuracy alone does not guarantee faithful reasoning.

5. Significance and Claims

The paper claims that MR-IQA-2 moves Blind Image Quality Assessment beyond "plausible explanations" toward visually verifiable understanding.

• Faithfulness over Fluency: The work demonstrates that generating fluent text is insufficient for IQA; the reasoning must causally link to the visual content. By using visual intervention as a test, the framework ensures that identified quality factors are genuine constraints on image quality.
• Downstream Utility: The faithful visual understanding acquired by the framework is not limited to scoring; it can inform image-quality optimization and related downstream tasks (e.g., image editing, visual recognition) by identifying specific, actionable quality-limiting factors.
• Methodological Shift: The paper positions the decoupling of reasoning and rating supervision as a critical step for RL-based BIQA, addressing the "reward ambiguity" that has plagued previous reinforcement learning approaches in this domain.

The authors acknowledge limitations, noting that the current framework does not yet form a fully closed-loop reflection (the edited image is not fed back to the Actor for further reasoning) and that the separation of Actor, Editor, and Judge introduces computational redundancy. However, they assert that the proposed mechanism provides a robust, verifiable path toward modeling human visual preferences and aesthetic judgments.