acceptodds
Under review as a conference paper at ICLR 2027

ROSA-IQA: Learning with Reasoning for Image Quality Assessment via On-Policy Self-Distillation

Abstract

Image quality assessment (IQA) takes two complementary forms: quality scoring quantifies perceptual quality, while quality description reveals the attributes and defects underlying human judgments. Although multimodal large language models (MLLMs) show promising reasoning capabilities, adapting them for quality description remains constrained by the fact that widely used IQA datasets typically provide only a scalar score per image. On-policy reinforcement learning directly learns from quality scores but provides only coarse sequence-level credit, whereas off-policy dense distillation from external MLLMs introduces a distribution shift between teacher demonstrations and student rollouts. We introduce a Reasoning-based On-policy Self-distillation Approach for IQA (ROSA-IQA). Under a teacher-student distillation paradigm, ROSA-IQA uses a frozen snapshot of the MLLM being optimized as the self-teacher, conditioned on curated privileged quality scores and segmentation-derived region guidance, to provide token-level supervision for student-sampled reasoning rollouts. We further use factual–counterfactual disagreement to identify tokens sensitive to quality conditions, thereby emphasizing quality-relevant descriptions in the student's reasoning. A scoring head then maps the combined visual and textual representations to continuous quality predictions. Extensive experiments show that ROSA-IQA outperforms leading no-reference IQA (NR-IQA) models in quality scoring while generating human-aligned quality descriptions. Importantly, we show that learning with reasoning improves cross-dataset generalization, suggesting that it acts as a regularizer that encourages the model to learn transferable quality features.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Quality Assessment, On-Policy Self-Distillation, Image Quality Reasoning, Privileged Information

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_B4QYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ROSA-IQA adapts on-policy self-distillation to learn image-quality descriptions and scores from mean opinion score (MOS) annotations without human-written descriptions. A frozen teacher initialized from the student backbone receives privileged MOS and segmentation-derived crops, and supervises next-token distributions along student-generated trajectories. Disagreement between teacher distributions under factual and counterfactual MOS selects positions for distillation. A continuous scoring head is trained jointly. Using Qwen3-VL-4B trained on KonIQ, the method achieves the highest reported average scoring performance across seven IQA datasets and favorable description evaluations. Direct scoring slightly outperforms reasoning-then-scoring at inference, which the authors interpret as evidence that description supervision acts as a training regularizer.

Critical strengths
  1. A useful objective and coherent implementation. Learning quality descriptions from existing score annotations reduces dependence on human-written explanations. The pipeline is clearly described, and scoring experiments cover authentic, synthetic, algorithm-induced, and AI-generated image distortions.
  2. Informative controlled comparisons. Table 3 examines score-only training, off-policy distillation, region guidance, and equal-budget random token selection. Table 4 also shows that direct scoring retains the benefits of joint training without generating descriptions at inference.
Critical weaknesses
  1. The motivation does not establish a clear bottleneck that the proposed extensions resolve. The paper invokes annotation and training costs, exposure bias, and coarse RL feedback to motivate its adaptation of OPSD. However, the experiments do not clearly identify which limitation constrains existing IQA performance. The on-/off-policy comparison shows a performance difference without establishing exposure bias as its cause, and cost savings are not quantified. This leaves the main additions (MOS/region conditioning and counterfactual token selection) without a strong account of why they are needed. For token selection in particular, comparisons with full-token and random controls support its usefulness but provide limited evidence for the specific selection criterion's advantage over other baseline methods.
  2. The demonstrated benefits are modest relative to the methodological claims. Since the method learns both scores and descriptions, either output could establish its practical contribution. For scoring, Table 1 improves average SRCC/PLCC over same-backbone IQA-T1 by 0.003/0.014. Table 3's gains over score-only training are concentrated in CSIQ, and token selection adds 0.004/0.003 over full-token distillation with the same region guidance. Description results provide some complementary evidence, but the largest gains concern a confidence criterion defined independently of correctness. Human preferences cover only 12 selected images per dataset across three datasets. Together, these results demonstrate benefits in selected settings without establishing a substantial, broadly supported advance.
Questions for authors
  1. Which limitation of existing IQA methods is principally responsible for the gains? What evidence distinguishes the proposed explanation from a general benefit of additional description supervision, and supports the particular counterfactual token-selection criterion?
  2. If reduced cost is a principal benefit, what are the training and direct-scoring inference costs relative to relevant baselines, particularly IQA-T1? Please include teacher computation and region processing.
3 comments
AuthorsOpenReview ↗

Looking Forward to Further Discussion

Dear Reviewer,

Thank you for your constructive feedback. We have added experiments and clarifications addressing your concerns, responded in detail in our point-by-point reply, and uploaded the revised manuscript. We are eager to hear any further suggestions you may have and hope that the revisions meet with your approval.

We clarified the role of each component. MOS supervises image-level scoring, SAM-2 regions give the teacher local distortion cues, and differences between the teacher distributions under factual and counterfactual MOS guide token selection. We present exposure bias as a motivation for on-policy training and report the on-policy and off-policy results as an empirical comparison. To examine the roles of additional description supervision and quality-sensitive token selection (QS), we compared offline quality-description distillation, generic image-description distillation, and feature distillation. We also added a token analysis, blind assessments of distortion type and severity, and a separate evaluation of region guidance for localization.

We clarified the size and dataset dependence of the scoring gains, added description baselines, and expanded the blinded human preference study to five methods and 150 images. ROSA-IQA received 41.3% of the votes on images selected by description disagreement. We also compared training costs with GRPO, including teacher computation, region encoding, and SAM-2 preprocessing, and clarified how our training procedure differs from IQA-T1's. We reported latency for both description generation and direct scoring. ROSA-IQA takes an average of 0.0378 seconds per image to score directly.

Thank you for the additional experiments and clarifications. The expanded description evaluations and cost analysis strengthen the empirical evidence. However, the scoring gains remain modest and uneven, and the new results do not sufficiently change my assessment of the significance of the methodological contribution. While several concerns have been alleviated, my overall judgment remains unchanged, and I maintain my original rating.

Thank you for taking the time to reply again! We regret that the scoring gains of our model did not fully meet your expectations. However, we would like to note that achieving competitive SOTA performance with a single forward pass, without requiring iterative tool calls or multiple forward passes at inference, is already noteworthy. Comparing our OPSD-based framework with GRPO-only methods (such as Q-Insight and VisualQuality-R1) on score regression may also provide a fairer perspective. Moreover, scoring directly on latent features enables continuous score modeling and fundamentally avoids format or parsing failures commonly encountered in discrete token-level score regression.

On the other hand, generating accurate and reliable quality descriptions is an equally central contribution of this work. In our LLM-as-a-Judge evaluation, ROSA-IQA achieves mean accuracy scores of 5.782 / 6.445 under Gemini 3.8 Flash and GPT 6.1 Sol, outperforming Q-Insight (5.532 / 6.198) and VisualQuality-R1 (5.440 / 6.273). In subjective study, ROSA-IQA attains an overall selection rate of 41.3%, substantially ahead of Q-Insight (16.9%) and VisualQuality-R1 (12.5%).

Finally, we kindly hope you will take into account the novelty of introducing this new training paradigm to IQA, as OPSD offers promising avenues for further meaningful research in privileged-information conditioning and selective token distillation. Thank you again for your thorough review and valuable feedback!

Reviewer_uAbyOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ROSA-IQA, a framework for image quality assessment based on on-policy self-distillation. The method uses a frozen self-teacher conditioned on privileged information, including ground-truth MOS and SAM-2-generated image regions, to supervise student-generated quality descriptions. It further introduces a factual–counterfactual disagreement mechanism to identify quality-sensitive tokens for selective distillation. A separate scoring head is trained to predict continuous quality scores from the learned representations. Experiments across seven IQA datasets show competitive quality prediction performance, along with improvements in generated quality descriptions. The authors also provide ablation studies to examine the effects of self-distillation, privileged information, and token selection.

Critical strengths

The paper addresses a relevant problem in IQA, where numerical quality scores are widely available but detailed quality descriptions are expensive to annotate. Using privileged information to provide token-level supervision is a reasonable approach to reducing the dependence on human-written descriptions.

The experimental evaluation covers multiple distortion scenarios and includes comparisons with existing IQA methods, component ablations, and human preference evaluation. The results suggest that on-policy self-distillation can improve cross-dataset performance while preserving the ability to generate quality descriptions.

Critical weaknesses
  1. The methodological novelty seems somewhat limited. The overall framework largely follows existing on-policy self-distillation approaches, while the main additions are MOS-conditioned supervision, semantic region guidance, and quality-sensitive token selection. Although these components are reasonable, the paper does not clearly establish what is fundamentally new compared with existing privileged-context and token-selective distillation methods.

  2. I am not fully convinced that the proposed factual–counterfactual disagreement accurately identifies visually meaningful quality-related tokens. Changing the MOS condition may simply cause the teacher to favor different quality-related expressions, without necessarily improving the grounding of those expressions in actual image distortions. The paper mainly provides qualitative examples but lacks a direct evaluation of whether the selected tokens correspond to correct perceptual evidence.

  3. The improvements from the proposed quality-sensitive distillation are relatively small and inconsistent across datasets. Compared with full-token distillation, the average SRCC increases only from 0.783 to 0.787, and the method performs worse on several individual datasets. Much of the improvement appears to come from CSIQ. It is therefore unclear whether the proposed token selection provides a consistent advantage across different distortion distributions.

  4. The contribution of semantic region guidance is not sufficiently demonstrated. Adding SAM-2 regions produces only marginal improvements in the ablation study, and the paper does not compare semantic regions with random crops or other region selection strategies. Since semantic segmentation does not necessarily identify perceptually important distortions, it remains unclear whether SAM-2 provides any particular advantage for IQA.

  5. The role of explicit quality reasoning in improving scoring performance remains unclear. Interestingly, Table 4 shows that direct scoring slightly outperforms reasoning-based scoring on average. While this does not rule out the benefits of reasoning supervision during training, the current experiments do not sufficiently distinguish these benefits from those of general representation distillation. In addition, the human preference evaluation uses a small set of deliberately selected description pairs, which limits how broadly the reported preference results can be interpreted.

Questions for authors
  1. What is the main methodological contribution beyond applying existing on-policy self-distillation with privileged information to IQA? How does the proposed token selection differ substantially from existing token-importance or selective distillation methods?

  2. How do the authors verify that tokens selected through factual–counterfactual disagreement are genuinely related to visual quality rather than simply being sensitive to the provided MOS condition? Could the authors provide a more direct evaluation of their perceptual relevance?

  3. The quality-sensitive distillation mechanism provides relatively small average improvements over full-token distillation and underperforms it on several datasets. Why are the gains concentrated in certain datasets, particularly CSIQ? Does the method behave differently across distortion types?

  4. What is the motivation for using SAM-2 semantic regions rather than random crops or distortion-aware regions? Can the authors provide evidence that semantic region selection is responsible for the observed improvements?

  5. Since direct scoring achieves better average performance than reasoning-based scoring, what evidence supports the claim that quality reasoning itself contributes to improved representations? Could similar improvements be obtained through additional distillation supervision without generating quality descriptions?

3 comments
AuthorsOpenReview ↗

Looking Forward to Further Discussion

Dear Reviewer,

Thank you for your constructive feedback. We have added experiments and clarifications addressing your concerns, responded in detail in our point-by-point reply, and uploaded the revised manuscript. We are eager to hear any further suggestions you may have and hope that the revisions meet with your approval.

We clarified how ROSA-IQA uses semantic region guidance and token selection based on factual and counterfactual MOS to jointly learn image quality scoring and description generation through on-policy self-distillation. To assess the perceptual relevance of selected tokens, we added an analysis of their semantic roles and blind assessments that used the images as the primary evidence for distortion type and severity. We added experiments on TID2013 and LIVE IQA. Quality-sensitive token selection (QS) improved both SRCC and PLCC on these two datasets as well as on KADID and CSIQ. We discussed how the gains vary between controlled synthetic distortions and authentic images with multiple interacting distortions.

To assess the contribution of SAM-2, we added a localization evaluation and a random-crop control matched in both the number and sizes of crops. To examine the role of quality-description supervision, we compared offline quality-description distillation, generic image-description distillation, and feature distillation. We also expanded the blinded human preference study to five methods and 150 images selected by description disagreement, collecting 1,050 judgments, and added a GRPO training cost comparison that accounts for teacher computation, region encoding, and SAM-2 preprocessing.

I appreciate the authors' detailed response. However, in my view, this work still largely represents an adaptation of OPSD to the IQA domain, primarily reusing many of the ideas and techniques already established in OPSD. Therefore, I believe that the paper offers limited contributions and insights. I maintain my original score.

AuthorsOpenReview ↗

Further Clarification

Thank you for your reply! We would like to make one brief final clarification on our contributions. As the first work to introduce OPSD into IQA, ROSA-IQA brings a more efficient training paradigm to GRPO-dominated MLLM-IQA that can more flexibly leverage auxiliary information. While privileged information and token selection exist in prior OPSD studies, our SAM region guidance for local distortion cues and our quality-sensitive token selection (QS) are specifically designed for IQA, where only scalar MOS labels are available without ground-truth descriptions. Their gains in both description accuracy and scoring performance are consistently verified by our ablations, LLM-as-a-Judge evaluations, and human studies. We kindly hope you can further consider the novelty of our work and its potential value to the community!

Reviewer_3NHhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper jointly learns quality description and scoring from IQA data with only MOS labels. Based on On-Policy Self-Distillation (OPSD), a frozen copy of the model serves as the teacher, receiving MOS and SAM-2 regions as privileged information. Along student-generated descriptions, token-level supervision is applied only at the 15% of positions with the largest teacher-distribution differences under factual and counterfactual MOS conditions. An MLP predicts a continuous quality score from the student's last-token hidden state, which encodes both the image and its description. Experiments report the best average SRCC/PLCC among compared methods across seven datasets.

Critical strengths

S1. The supervision strategy shows practical value. By providing MOS and local regions as privileged information, the teacher offers token-level supervision for student-generated descriptions. This enables description learning from existing IQA data without description annotations and reduces reliance on external MLLMs to generate training texts. S2. The counterfactual MOS-based selection makes OPSD more task-specific to IQA. The method selects supervision positions based on teacher-distribution differences under factual and counterfactual conditions. It achieves higher average SRCC/PLCC than both full-token supervision and random 15% selection, suggesting that the gains come from informative selection rather than sparsity alone.

Critical weaknesses
  1. Shifting MOS by 50 while keeping the image unchanged identifies score-sensitive positions, which may reflect evaluative wording rather than actual defects. A large distributional difference may simply mean that the teacher follows the supplied score-for example, favoring "sharp" under a high MOS and "blurry" under a low MOS-rather than identifying defects from the image. The scoring gains support the selection strategy, but do not show that the selected positions provide more accurate visual quality evidence.

  2. The quality-description evaluation in this paper is limited to Q-Insight. VisualQuality-R1 also supports quality descriptions, while Zoom-IQA and IQA-T1 provide region- or evidence-based reasoning. All three appear in Table 1, yet only the scoring performance is evaluated. Since learning accurate quality descriptions is a central motivation in this paper, comparison with Q-Insight alone provides limited support for the claimed descriptive advantage.

Questions for authors
  1. Figure 4 shows only one selection example. Can the authors provide correctness analyses and failure cases? Why is Δ=50\Delta = 50 chosen, and are the results stable across offsets?

  2. Can the authors compare description accuracy with and without region guidance and quality-sensitive selection under the same evaluation protocol?

  3. The introduction motivates OPSD through the cost of multiple RLVR rollouts, but no training-cost comparison is reported in this paper. Does it remain cheaper after counting both teacher evaluations, region encoding, and SAM-2 preprocessing? Can the authors report training time, GPU-hours, or peak memory on matched hardware.

  4. RALI [1] is neither discussed nor compared. RALI already studies quality descriptions as cross-domain representations and enables scoring without reasoning text at inference. These directly relate to the paper's discussion of description learning, cross-dataset generalization, and direct scoring. The paper should clarify its contribution beyond these findings.

  5. The discussion of related work on OPSD is incomplete. Vision-OPD [2] already uses a crop-conditioned teacher to supervise a full-image student along student-generated trajectories. ViCuR [3] and ViGOS [4] further examine whether privileged teachers base their supervision on visual evidence or rely mainly on the supplied answers. None is discussed, although they directly relate to the proposed region guidance and MOS-conditioned supervision.

[1] Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality Assessment, 2026. [2] Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation, 2026. [3] ViCuR: Visual Cues as Recoverable Privilege for Multimodal On-Policy Distillation, 2026. [4] Seeing Before Reasoning: Decoupling Perception and Reasoning for Shortcut-Resilient Multimodal On-Policy Self-Distillation, 2026.

3 comments
AuthorsOpenReview ↗

Looking Forward to Further Discussion

Dear Reviewer,

Thank you for your constructive feedback. We have added experiments and clarifications addressing your concerns, responded in detail in our point-by-point reply, and uploaded the revised manuscript. We are eager to hear any further suggestions you may have and hope that the revisions meet with your approval.

To assess the visual relevance of MOS-sensitive tokens, we analyzed their semantic roles and asked Gemini 3.8 Flash and GPT 6.1 Sol to conduct blind pairwise comparisons of the descriptions. Quality-sensitive token selection (QS) more often selected tokens that express distortion severity or the direction of a quality judgment. The resulting descriptions also characterized distortion type and severity more accurately. We compared MOS offsets of Δ = 20, 40, and 50 and explained why we used Δ = 50. We added a distortion localization evaluation to assess region guidance and included VisualQuality-R1, Zoom-IQA, and IQA-T1 in our description evaluation.

We compared training costs with GRPO on the same hardware and under matched training settings. Including both teacher forward passes, region encoding, and one-time SAM-2 preprocessing, the estimated total time was about 25% of the GRPO training time. We also added RALI to the main results table, clarified how the two methods differ in training supervision and direct scoring, and expanded the related work discussion to cover Vision-OPD, ViCuR, and ViGOS.

Thanks for the response and additional experiments. They address several of my concerns, but I still have doubts about the core method. Changing the MOS may simply change the teacher's wording, but this doesn't necessarily mean the selected tokens better reflect the actual image quality. Also, the improvement over simpler distillation strategies is quite limited. Overall, I still feel the technical novelty is limited, so I will keep my rating unchanged.

AuthorsOpenReview ↗

Further explanation

Thank you for your prompt and thoughtful response!

We would also like to offer one additional explanation. In fact, determining which tokens more accurately reflect the true perceptual quality of an image is inherently difficult, both from the model's perspective and in terms of human perception. The QS mechanism does not attempt to solve this problem. Instead, it identifies tokens that are most sensitive to changes in quality scores and thus require greater alignment with human subjective ratings, making them the primary targets for correction during distillation. Therefore, our contribution lies not only in introducing OPSD to IQA for the first time, but also in designing a more targeted distillation mechanism that better aligns the model's quality perception with human subjective judgments.

Experimentally, introducing QS leads to average improvements of 1.79% and 1.58% in SRCC and PLCC, respectively, across four synthetic distortion datasets that are sensitive to distortion severity, while maintaining comparable performance on other distortion datasets. We believe this represents a meaningful improvement.

We hope these clarifications and experimental results will help you further consider the novelty of our work and the potential value of this new training paradigm for IQA. Of course, we fully respect your final decision, and we greatly appreciate the time and effort you have devoted to reviewing our work. Thank you again for your valuable feedback!

Comments on the paper

1 comment
Area_Chair_ScqYOpenReview ↗

Reviewer-author discussion period is ending. Please feedback.

Dear Reviewers,

Thanks for your time and effort in reviewing ICLR2027 submissions.

As the reviewer-author discussion period will end very soon, please take time to read the author’s rebuttal and give your feedback. Your continuous engagement is highly appreciated.

Best regards,

Your AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.