acceptodds
Under review as a conference paper at ICLR 2027

AesMind: Harnessing Multimodal Large Language Models for Neuroaesthetic Understanding and Assessmen

Abstract

Recent advances in multimodal large language models (MLLMs) have significantly enhanced the interpretability and generalization of image aesthetic assessment (IAA). However, existing methods still predominantly evaluate aesthetics through either a single holistic score or a generic critique, failing to provide fine-grained attribute ratings and structured multidimensional reasoning. To address these limitations, we draw upon neuroaesthetics, which explores the cognitive and neural mechanisms of aesthetic appreciation, to establish a comprehensive aesthetic experience taxonomy encompassing the sensory-motor, emotion-valuation, and knowledge-meaning systems. Based on this theoretical framework, we first introduce NeuroAes, a large-scale dataset specifically tailored for computational photographic NeuroAesthetics, which comprises 25K images, annotated with over 1.2M fine-grained neuroaesthetic labels across 48 subcategories and over 2.25M human ratings spanning the valence-arousal-dominance (VAD) dimensions for both aesthetic and evoked emotions. In particular, NeuroAes is partitioned into a 5K-image NeuroAes-Bench for comprehensive aesthetic evaluation and a 20K-image NeuroAes-Data for large-scale training. Building upon this foundation, we develop AesMind, an all-in-one cognitive aesthetic MLLM that unifies perception, assessment, and description capabilities through a progressive three-stage training paradigm of skill harnessing, supervised fine-tuning (SFT), and group relative policy optimization (GRPO). Extensive experimental results demonstrate that AesMind achieves state-of-the-art performance across all benchmark tasks. Notably, it exhibits superior continuous-emotion regression, unprecedented cross-dataset generalization, and the unique capability to give aesthetic improvement suggestions, proving its ability to deeply understand photographic aesthetics rather than merely reciting superficial visual cues. Both NeuroAes and AesMind will be released to facilitate further research.

Primary area:
datasets and benchmarks
Keywords:
Image Aesthetics Assessment, Multimodal Large Language Models, Neuroaesthetics

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_D7zvOpenReview ↗
Rating
4(+1)
Confidence
4(+1)
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper introduces NeuroAes, a neuroaesthetics-grounded dataset and benchmark for photographic aesthetic understanding, together with AesMind, a specialized multimodal large language model. NeuroAes contains 25K images with fine-grained annotations spanning sensory-motor attributes, aesthetic and evoked emotions, and structured professional critiques, including human-rated continuous VAD scores. Based on this dataset, AesMind is trained through a three-stage pipeline consisting of skill harnessing, supervised fine-tuning, and GRPO with task-specific rewards. Experiments show strong performance on the proposed benchmark and competitive cross-dataset generalization, particularly for fine-grained perception, continuous aesthetic assessment, and structured photographic critique generation.

Critical strengths
  1. The paper presents a solid and well-motivated neuroaesthetics-grounded taxonomy that goes beyond a single aesthetic score and organizes photographic assessment into perception, emotion, and critique.

  2. NeuroAes is a substantial and carefully constructed dataset, with 25K images, fine-grained labels, human-rated VAD annotations, and a dedicated 5K-image benchmark covering perception, assessment, and description tasks.

  3. The experimental evaluation is comprehensive and generally convincing, including strong results on the proposed benchmark, ablations of the three training stages, and cross-dataset evaluation on several existing aesthetic benchmarks.

Critical weaknesses
  1. The annotation pipeline mixes human judgments with Gemini-generated annotations, but the paper does not provide sufficient calibration between these two sources. It would be useful to report how much of the training data is purely human-annotated, how well Gemini annotations agree with human judgments, and ablations comparing training with human-only, LLM-only, and combined annotations.
Questions for authors
  1. Could the authors provide more details about the human annotators’ backgrounds, especially the distribution of their artistic and photographic expertise? I am particularly interested in whether these annotators can truly be considered experts, given that the paper emphasizes professional-level aesthetic assessment.

  2. Since much of the evaluation is conducted within the NeuroAes taxonomy and annotation framework, I am curious how AesMind would perform on independent expert-grounded benchmarks. In particular, Visual Aesthetic Benchmark [1] is aligned with expert consensus on aesthetic preferences, and AesBench [2] provides another external benchmark for expert-level aesthetic assessment. Results on these benchmarks would make the generalization claims more convincing.

  3. The annotation pipeline combines human judgments and Gemini-generated annotations, but it is unclear how well these two sources are calibrated. Could the authors report how much of the training data is purely human-annotated, the agreement between human and Gemini annotations, and ablations comparing training on human-only, LLM-only, and combined annotations?

[1] https://arxiv.org/abs/2605.12684 [2] https://arxiv.org/abs/2401.08276

6 comments

To Reviewer D7zv:

Thank you for the constructive and detailed feedback. We address these points below and will incorporate the corresponding analyses and clarifications into the revised manuscript.

1. W1 & Q3: Annotation source and human calibration

Thank you for highlighting the need to clarify annotation source and quantify the role of human verification. NeuroAes contains three complementary annotation forms corresponding to the Assessment, Perception, and Description tasks in NeuroAes-Bench. Human judgments and Gemini outputs play different roles in their construction, rather than being mixed together as interchangeable labels pooled for the same task.

(a) Annotation sources and training-data composition

As illustrated in Step 2 of Figure 2, three tasks have different annotation processes. (1) For assessment task, the six continuous VAD dimensions—valence, arousal, and dominance for aesthetic and evoked emotions—are entirely human-rated. Each image is rated by 15 participants, and the ratings are standardized and aggregated into MOS. Gemini does not generate or replace these scores. (2) For perception task, Gemini produces preliminary annotations that are subsequently checked and corrected by human annotators. Specially, labels of the emotion-valuation system are additionally grounded in the human VAD scores. (3) For description task, the existing human-written comments in the image metadata serve as the primary reference, with the image and human-verified labels providing supplementary grounding for Gemini's structured generation. These critiques are therefore human-grounded but Gemini-generated, not independently human-authored. The annotation types, data quantities, and corresponding sources for the three tasks are summarized in the table below and in Table 15 in the revised paper.

TaskAnnotationQuantitySource
AssessmentIndividual VAD ratings (6 dimensions)2.25M ratings15 human raters per image
AssessmentVAD mean opinion scores (MOS)150K scoresAggregated from human ratings
PerceptionMotor-Sensory labels25K records (~1M labels)Gemini annotation with human verification
PerceptionEmotion-Valuation labels25K setsGemini annotation grounded in images and human VAD scores
DescriptionAesthetic critiques (8 dimensions)25K recordsGemini generation based on images, human-written comments, and verified labels

As shown in the table below and in Table 16 in the revised paper, 11.75% of the training QA examples are purely human-annotated, with continuous VAD targets derived exclusively from human ratings. The remaining 88.25% involve human contributions through verification, correction, or reference annotations, combined with Gemini-generated content.

TaskHuman AnnotationTraining QA ExamplesPercentage
AssessmentPure120,00011.75%
PerceptionPartial752,63673.70%
DescriptionPartial148,61014.55%
Total--1,021,246100%

1. W1 & Q3: Annotation source and human calibration (continued)

(b) Human verification and Gemini annotation quality

To assess the quality of Gemini-generated annotations, we analyzed the human-review logs, which reveal varying levels of human intervention across annotation dimensions. Across 41 fields, we identified 77,476 unique image–field records with logged human interventions among 971,305 post-review entries, corresponding to an overall intervention rate of 7.976%. The following table (Table 13 in the revised paper) illustrates substantial differences in the extent of human review across annotation dimensions, particularly between fine-grained photographic and affective attributes and basic semantic attributes.

Annotation fieldFormatHuman-review recordsPost-review entriesIntervention rate
Additional composition patternsMultiple-choice18,22823,73176.811%
Portrait compositionMultiple-choice6,79512,29855.253%
Color-quality evaluationSingle-choice11,16925,00044.676%
Portrait angleMultiple-choice3,77312,29130.697%
Discrete emotion labelSingle-choice6,37825,00025.512%
Number of peopleSingle-choice15725,0000.628%
Main subjectOpen-ended16425,0000.656%
Animal presenceSingle-choice2225,0000.088%

The results reveal clear dimension-wise differences in Gemini annotation quality, as reflected by the extent of subsequent human intervention. Even among single-choice fields, color-quality evaluation and discrete emotion labels exhibit substantially higher intervention rates than basic semantic attributes such as people counting and animal presence. This suggests that Gemini provides more reliable preliminary annotations for basic semantic recognition, whereas fine-grained aesthetic and affective attributes require substantially more human verification. The high intervention rates for composition patterns and portrait framing further highlight the difficulty of accurately annotating complex photographic properties. Although intervention rates may also be influenced by annotation format and field applicability, they provide empirical evidence that human verification plays a particularly important role in refining Gemini-generated annotations for subjective and fine-grained aesthetic dimensions.

(c) Human-only, LLM-only, and combined training

Thank you for raising this important question regarding the contribution of different annotation sources to model training. A controlled comparison of human-only, LLM-only, and combined training is challenging due to the task-specific annotation protocols and the resulting lack of comparable training targets across sources.

Specifically, for Perception and Description, annotations are initially generated by Gemini and subsequently refined or verified by human annotators. Thus, independently human-authored annotations are not available for these tasks. Moreover, directly using raw Gemini-generated annotations without human refinement would introduce substantial annotation errors, making it difficult to establish a meaningful LLM-only training baseline. Conversely, Assessment relies entirely on human ratings to derive continuous VAD targets, and no corresponding LLM-generated targets are available for an LLM-only comparison. Therefore, the three training configurations cannot be consistently constructed across all tasks without changing the task composition and supervision targets.

2. Q1: Participant backgrounds and annotation training

Thank you for asking for a clearer account of the annotators' artistic and photographic experience. All 71 participants took part in the subjective rating experiment, and we collected background questionnaires and Big Five personality measures from each of them. We additionally invited a subset with photographic experience to perform human label verification. The table below (Table 11 in the revised paper) summarizes the full rating cohort, rather than the verification subset alone.

(a) Background composition

BackgroundCategoryNPercentage
GenderMale3549.3%
GenderFemale3650.7%
EducationUndergraduate2636.6%
EducationPostgraduate4563.4%
Artistic experienceNovice2940.8%
Artistic experienceIntermediate3042.3%
Artistic experienceAdvanced1216.9%
Photographic experienceNovice2433.8%
Photographic experienceIntermediate3650.7%
Photographic experienceAdvanced1115.5%

Across the rating cohort, 66.2% reported intermediate or advanced photographic experience, and 59.2% reported intermediate or advanced artistic experience. These are questionnaire-based experience levels rather than verified professional credentials. For label verification, participants were selected from those with photographic experience. The background and personality questionnaires document participant variation, rather than serving as a substitute for task-specific preparation.

(b) Standardized training and field-specific guidance

All 71 rating participants received standardized training, irrespective of their prior experience. The verification subset additionally received a detailed manual for each annotation field. These manuals define the dimension and its response options, illustrate the options with multiple photographic examples, and explain ambiguous or difficult cases. This combines the verifiers' photographic experience with a common operational basis for checking labels. As reported in our first response, the review logs further document substantial human intervention in fields such as color-quality evaluation and portrait composition.

In summary, the rating cohort includes a majority with photographic experience, while label verification is performed by a selected photography-experienced subset supported by standardized training and detailed field-specific guidance. We have included the participant backgrounds and Big Five summary statistics in the supplementary material, with the updates highlighted in teal.

Reviewer_5B6GOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces NeuroAes, a 25K-image neuroaesthetics-inspired dataset and benchmark covering fine-grained perception, continuous VAD assessment, and structured aesthetic critique. Based on NeuroAes, the authors develop AesMind, which combines skill harnessing, LoRA-based SFT, and GRPO with task-specific rewards, and reports strong performance on NeuroAes-Bench and several external aesthetics benchmarks.

Critical strengths
  1. Comprehensive dataset and taxonomy. NeuroAes provides unusually fine-grained annotations spanning perceptual attributes, aesthetic/evoked emotions, and professional critiques.
  2. The paper evaluates multiple capabilities rather than reducing aesthetic assessment to a single MOS score.
  3. The empirical results are strong, particularly for continuous VAD prediction and fine-grained perception. For example, AesMind substantially outperforms its Qwen3.5-9B backbone on NeuroAes-Bench.
Critical weaknesses
  1. Limited methodological novelty. The core pipeline is largely a combination of existing techniques: skill/prompt optimization, LoRA-based SFT, and GRPO. The proposed hybrid reward is relatively heuristic, making the technical contribution less compelling than the dataset contribution.
  2. The baseline comparison is not sufficiently strong or current. For specialized image-aesthetic models, the comparison is dominated by 2024-2025 approaches such as Q-Align, UniAA, Next Token Is Enough, and ArtiMuse. More recent and stronger domain-specific baselines would make the claimed state-of-the-art result more convincing.
  3. The ablation suggests a non-trivial trade-off: adding GRPO substantially improves continuous assessment but actually reduces overall perception accuracy compared with Skill+SFT (70.53 → 65.36). This weakens the claim that the complete three-stage pipeline is uniformly beneficial.
Questions for authors
  1. What is the key technical novelty of AesMind beyond combining skill harnessing, SFT, and GRPO with manually designed rewards?
  2. Could the authors include stronger and more recent specialized aesthetic-assessment baselines, particularly methods from 2025–2026?
  3. Why does GRPO noticeably reduce perception accuracy compared with Skill+SFT, and how should this trade-off be interpreted?
7 comments

To Reviewer 5B6G:

Thank you for the constructive and detailed feedback. We address these points below and will incorporate the corresponding analyses and clarifications into the revised manuscript.

1. W1 & Q1: Technical contribution beyond combining existing techniques

Thank you for the thoughtful feedback. We clarify the relative contributions of NeuroAes and AesMind, and provide evidence addressing the concerns about component integration and reward design.

(a) Dataset and methodological contributions

A key contribution of this work is the construction of the NeuroAes dataset, and we appreciate your positive assessment of its fine-grained taxonomy and multi-capability evaluation. NeuroAes connects perception, emotion, and professional critique across 25K photographs, with over 1M fine-grained labels and continuous VAD annotations derived from over 2.25M human ratings. In particular, it distinguishes aesthetic emotions associated with photographic execution from emotions evoked by image content, supporting evaluation beyond a single aesthetic score. AesMind complements this resource by exploring a training recipe that supports perception, assessment, and description within one unified model.

(b) Skill integration beyond component stacking

While skill harnessing, LoRA-based SFT, and GRPO are established techniques, the technical contribution lies in how prompt-level skill harnessing interacts with parameter-level post-training, rather than in introducing these components individually. To our knowledge, the joint use of these techniques for aesthetic understanding and assessment has not been explored previously. More importantly, their combination is not interchangeable, e.g., skills that are effective when introduced prior to training may not necessarily benefit a model that has already undergone specialized training. Before finalizing AesMind, we explored alternative integration orders, including attaching the base-model skill after SFT or SFT + GRPO. We report these earlier results alongside the training-stage ablations in the table below and in Table 20 in the revised paper :

ModelSkill placementPerception (ACC)Assessment (SRCC)Assessment (PLCC)
Qwen3.5-9BNone45.650.25370.2613
Qwen3.5-9B + SkillInference only46.950.44920.4956
Qwen3.5-9B + SFTNone62.230.48710.5270
Qwen3.5-9B + SFT + SkillAttached after SFT70.530.44560.4941
Qwen3.5-9B + SFT + GRPONone66.210.65320.6707
Qwen3.5-9B + SFT + GRPO + SkillAttached after SFT + GRPO65.090.61690.6379
Qwen3.5-9B + Skill + SFTBefore SFT70.530.50960.5420
AesMind (Skill + SFT + GRPO)Before SFT (ours)65.360.69860.7261

The results show that skill placement matters most clearly for continuous assessment. Attaching the skill after specialization degrades assessment, even though it can improve perception after SFT; incorporating it before training improves assessment with both SFT and SFT + GRPO. During the same exploration, we also optimized new skills directly on the trained SFT and SFT + GRPO models. However, neither of these approaches yielded a better candidate. These findings indicate that simply stacking the components together fails to replicate the advantages provided by our progressive design.

These observations are consistent with the distinct roles of the stages: the skill first aligns task semantics through label definitions, decision criteria, and scoring anchors; SFT internalizes domain knowledge under this shared context; GRPO refines the predictions. Introducing a skill only after specialization may instead create a mismatch with the learned instruction–response mapping, while leaving less scope for prompt-level improvement. The design motivations provide theoretical basis and the experiments provide empirical basis for our integration, beyond simple combination.

1. W1 & Q1: Technical contribution beyond combining existing techniques (continued)

(c) Reward design and controlled ablation

The motivation of the proposed reward design comes from addressing two challenges, i.e., distinguishing precise VAD predictions from merely close ones, and counteracting the dominance of frequent categories in photographic attributes. To test their contribution beyond standard GRPO, we have further trained an additional Stage3-reward variant with simple alternatives: an L1 reward for continuous scoring and a uniform binary accuracy reward for single-choice perception. The GRPO algorithm and its optimization procedure were kept identical to the original setting; only the task-specific reward functions were replaced. Specifically, the format reward, curriculum gate, training data, and all training hyperparameters remained unchanged. Assessment and Perception provide task-specific comparisons, while the joint replacement tests the reward design as a whole. Row numbers follow Table 7 in revised paper (Table 6 in raw submission) , and "new" denotes the added run.

No.SkillLoRAGRPO rewardMotor ACCEmotion ACCOverall ACCSRCCPLCCWord2Vec
(6)✓✓–76.3147.9570.530.50960.54200.9250
new✓✓L1 + uniform binary69.2337.7262.810.69060.70470.9254
(7)✓✓Ours (hybrid + prior-weighted)70.8943.7565.360.69860.72610.9262

As shown in the table above and in Table 22 in the revised paper, the proposed rewards outperform the simple alternatives on all reported metrics. Overall perception accuracy improves by 2.55 percentage points, with a 6.03-point improvement in Emotion accuracy; assessment SRCC/PLCC improve by 0.0080/0.0214. Since both variants use GRPO under the same training settings, these results support an empirical benefit from the task-specific reward design beyond the use of RL post-training alone. Skill + SFT still achieves the highest perception accuracy in this table, so the reward improvements mitigate, rather than eliminate, the perception–assessment trade-off discussed in our third response.

Because one GRPO run takes at least five days on 8 × A100 GPUs, we replaced both rewards in a single run rather than ablating them separately. We omit the LLM-as-a-Judge Description metric here because its extensive API evaluation could not be completed within the available time and budget. After completing this evaluation, we will incorporate the full ablation results, the skill-placement experiments, and the corresponding analysis into the main text and supplementary material.

2. W2 & Q2: More recent baselines and external evaluation

Thank you for pointing out that stronger and more recent specialized baselines would make the comparison more convincing. We have expanded the evaluation with two recent models, additional baseline adaptation, and an external photography-aesthetics benchmark.

(a) Recent specialized models

We added UniPercept [1], first released on December 25, 2025, and PortraitAes [2], first released on October 4, 2026. UniPercept is a subsequent work from the ArtiMuse team, extending perceptual understanding to aesthetics, quality, structure, and texture through domain-adaptive training and task-aligned RL. PortraitAes focuses on intent-conditioned, structured assessment of portrait photography rather than general-purpose image aesthetics. Nevertheless, its emphasis on photographic execution and intent makes it a relevant complementary baseline; we distinguish its specialized scope when interpreting cross-domain performance.

The following table summarizes the overall NeuroAes-Bench results, with ArtiMuse included as an existing specialized reference:

ModelPerception ACC (%)Assessment SRCCAssessment PLCC
ArtiMuse33.110.22340.2835
UniPercept41.840.32590.4127
PortraitAes45.780.29030.3355
AesMind65.360.69860.7261

These two new baseline models outperform ArtiMuse in terms of overall perception and assessment, strengthening the comparison beyond the originally included methods. AesMind retains the best results among these models on both tasks, so its advantage is not restricted to comparison with older specialized models. Their description results, including skill-assisted variants, are reported below.

(b) Adapted baselines

We have also expanded the comparison beyond the original checkpoints by adding skills to representative baselines, including LLaVA-OneVision and ArtiMuse, and further fine-tuning ArtiMuse on NeuroAes-Data. The table below (Table 1 and Table 23 in the revised paper) reports perception accuracy for the Motor and Emotion systems and the overall task.

ModelAdaptationMotor ACC (%)Emotion ACC (%)Overall ACC (%)
LLaVA-OneVision-1.5-8BNone41.7018.8135.80
LLaVA-OneVision-1.5-8B+ Skill45.5218.2939.97
ArtiMuseNone38.8513.7133.11
ArtiMuse+ Skill43.9417.7438.60
ArtiMuse+ Skill + NeuroAes-Data fine-tuning69.9242.7164.37
UniPerceptNone46.2724.5541.84
UniPercept+ Skill48.5326.0643.95
PortraitAesNone52.0021.5545.78
PortraitAes+ Skill52.7424.1746.90
Gemini-3-ProNone30.2910.9026.28
Gemini-3-Pro+ Skill61.7126.4154.48
AesMind+ Skill + NeuroAes-Data fine-tuning70.8943.7565.36

Skill-assisted adaptation improves overall perception across these baselines, while fine-tuning on NeuroAes-Data further strengthens ArtiMuse. AesMind achieves the best Motor, Emotion, and overall perception accuracy among the compared models, including the adapted baselines. For continuous assessment, ArtiMuse fine-tuned on NeuroAes-Data achieves 0.5332/0.5571 SRCC/PLCC, compared with AesMind's 0.6986/0.7261. Thus, the comparison now includes both recent specialized models and baselines adapted to our tasks.

We have additionally evaluated description using the four text-similarity metrics below, covering the new models and the adapted baselines (Table 24 in the revised paper):

Model / MethodSimCSEWord2VecROUGE-LMETEOR
Qwen3.5-9B0.82850.90350.13050.1750
LLaVA-OneVision-1.5-8B0.80870.89540.18950.1635
LLaVA-OneVision-1.5-8B + Skill0.82080.90230.16860.1688
Gemini-3-Pro0.86370.92360.15160.2115
Gemini-3.1-Pro + Skill0.84340.91240.16030.1670
ArtiMuse0.82450.90690.16880.1961
ArtiMuse + Skill0.83110.90940.16390.1898
ArtiMuse + Skill + NeuroAes-Data fine-tuning0.85040.91940.24860.2115
UniPercept0.83270.91010.17790.1980
UniPercept + Skill0.81510.89580.15990.1693
PortraitAes0.84860.91550.14550.2104
PortraitAes + Skill0.83970.91260.17780.1785
AesMind0.85820.92620.25590.2335

For description, AesMind achieves the best Word2Vec, ROUGE-L, and METEOR scores, while Gemini-3-Pro achieves the best SimCSE score. Skill alone has model- and metric-dependent effects, whereas ArtiMuse fine-tuned with the skill improves across all four metrics. These results extend the comparison to all three capabilities evaluated by NeuroAes-Bench.

Reviewer_J49WOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a neuroaesthetics-inspired framework and dataset (NeuroAes) for computational photographic understanding, organizing evaluation into sensory-motor perception, emotion-valuation assessment, and knowledge-meaning critique generation. In particular, the dataset collects human ratings on continuous Valence-Arousal-Dominance (VAD) scales to separate aesthetic emotions induced by photographic execution from general emotions evoked by image content, supplemented by LLM-assisted attribute tags and critiques across 25,000 images. Building on this dataset, the authors train AesMind on top of Qwen3.5-9B through a three-stage pipeline consisting of system prompt optimization (skill harnessing), supervised fine-tuning with LoRA, and group relative policy optimization (GRPO). Experiments on NeuroAes-Bench and external aesthetic benchmarks show that the resulting model improves attribute classification, continuous VAD regression, and structured critique generation over general-purpose and existing aesthetic MLLMs.

Critical strengths
  1. Well-motivated affective formulation and subjective data collection: Grounding photographic assessment in the neuroaesthetic triad—and specifically decoupling execution-driven aesthetic emotions from content-driven evoked emotions within a continuous VAD space—is conceptually sound and addresses a real limitation of single-score IAA datasets. Collecting 15 independent human ratings per image with disjoint annotator pools between training and test splits provides a solid empirical basis for the continuous assessment task.
  2. Comprehensive multi-task benchmark and downstream evaluation: The paper evaluates perception, continuous emotion regression, and multi-dimensional critique generation within a unified pipeline, and further tests cross-dataset generalization on five conventional IAA benchmarks (AVA, PARA, TAD66K, FLICKR-AES, ArtiMuse-10K) as well as CogIP-Bench.
Critical weaknesses
  1. Methodological novelty, dataset annotation validity, and evaluation circularity:

    • Combinatorial training pipeline without reward ablations: From a modeling perspective, AesMind mainly stitches together existing post-training recipes—prompt optimization via SkillOpt (Yang et al., 2026a), standard LoRA SFT, and GRPO. The primary technical claims in Stage 3 lie in the continuous hybrid regression reward (Eq. 3) and the prior-weighted categorical matching reward (Eq. 4), together with the step-wise format curriculum gate (Eq. 5 and Eq. 8). However, the paper provides no ablation comparing these custom reward formulations against standard alternatives (e.g., a simple L1/MSE reward or uniform binary accuracy reward), making it unclear how much of the gain comes from the proposed reward design versus standard RL post-training.
    • Contradictions in "human-verified" annotations and reliance on LLM-generated ground truth: While the continuous VAD scores come from human raters, Appendix D.3.3 and D.3.4 show that the discrete emotion labels, dominant aesthetic/emotional factors, and all 8-dimensional ground-truth critiques are generated by Gemini-3-Pro. More concerningly, in Figure 9 (page 42), which presents a flagship sample from NeuroAes, the "human-verified" Sensory-Motor tags directly contradict both the image and the Knowledge-Meaning critique: the tags label the image as "cool tone" with primary colors ["blue", "green", "gray"], "Number_of_People": "two" (["male", "female"]), and "Time_of_Day": "daytime", whereas the image and critique describe "a lone silhouette on a traditional town bridge at sunset" unified in "rich amber and golden-orange tones (High saturation warm colors)". If such basic factual contradictions remain in a highlighted showcase example after human verification, it raises questions about the overall label quality of NeuroAes-Bench.
    • Bias in Description evaluation (Eq. 6 & Appendix F.1.3): While the authors note in Appendix E.2.3 that they introduce a deterministic word-count ratio score S_conc (Eq. 6) because LLM judges are insensitive to conciseness, this creates an unfair structural advantage when combined with Stage 1 Skill Harnessing. Specifically, AesMind's optimized Description Skill prompt (Appendix F.1.3, page 32) explicitly instructs the model to "Output ONLY the commentary in a single paragraph of exactly 2-3 sentences (50-80 words)" and avoid bullet points or headings. In contrast, zero-shot baselines (such as Gemini-3-Pro and ArtiMuse in Figures 11–12) naturally generate detailed multi-paragraph or bulleted critiques and are therefore heavily penalized by both S_conc and lexical overlap metrics (ROUGE-L, METEOR) against Gemini-generated reference texts.
  2. Missing experimental controls, multi-task trade-offs, and numerical anomalies:

    • Unmatched baseline settings in Perception and Description: In Table 2 (Assessment), the authors fine-tune ArtiMuse and MUSIQ on NeuroAes-Data (marked with ⋆), which substantially boosts ArtiMuse from 0.2234 to 0.5332 SRCC. However, in Table 1 (Perception) and Table 3 (Description), no baseline is fine-tuned on NeuroAes-Data or evaluated with the optimized Skill prompts. Notably, Gemini-3-Pro—the exact model used to pre-annotate the Perception labels (Appendix D.3.1)—scores only 26.28% in Table 1 (far below Qwen3.5-9B at 42.05%). Looking at the Perception Choice Skill (Appendix F.1.3), NeuroAes-Bench requires strict output formatting (e.g., outputting every applicable option on a new line for multi-choice questions) and dataset-specific tie-breaking rules; without these instructions or fine-tuning, zero-shot models suffer severe penalties under the Guessing-Penalty metric (Eq. 2).
    • Performance degradation from GRPO in the ablation study (Table 6): Although the paper claims that the three stages progressively enhance the model, comparing Row (6) (Skill + LoRA) and Row (7) (Skill + LoRA + GRPO) in Table 6 shows that adding GRPO actually degrades Perception accuracy across all categories: Motor drops from 76.31% to 70.89%, Emotion drops from 47.95% to 43.75%, and Overall drops by 5.17 points from 70.53% to 65.36%. Even compared to LoRA + GRPO without Skill (Row 4: 66.21% Perception, 1.2083 Description LLM score), the full model (Row 7: 65.36% Perception, 1.1840 Description) is worse on both Perception and Description. Furthermore, on Description (Gemini-3.1 judge), applying Skill alone to the frozen base model (Row 5: 1.2328) outperforms all fine-tuned and GRPO models (Rows 4, 6, and 7). While Appendix G.1 briefly notes an "optimization trade-off" on general semantic categories (Table 14), this appears closely tied to the training design (Appendix E.3.2 & F.4): SFT trains LoRA without <think> tags, whereas GRPO merges LoRA and performs full-parameter fine-tuning only on 50K single-choice and score QAs (omitting multi-choice and description tasks).
    • Inconsistent baseline numbers and metric bounds across tables:
      1. In Table 1, the base model Qwen3.5-9B achieves 42.79 (Motor), 19.86 (Emotion), and 42.05 (Overall) on Perception, whereas in Table 6 Row (1), the exact same un-finetuned baseline is reported as 53.21 (Motor), 19.28 (Emotion), and 45.65 (Overall) (while its Assessment and Description numbers remain identical to Tables 2 and 3).
      2. In Appendix E.2.3 (Eq. 7), the LLM-as-a-Judge score is explicitly defined as S_desc = (1/2) * (1/4) * (S_comp + S_prec + S_rele + S_conc), where each sub-score is in {0, 1, 2}, which strictly bounds S_desc in [0, 1]. However, in Tables 3 and 6, multiple models achieve scores above 1.0 (e.g., 1.1043, 1.1840, and 1.2328).
      3. In Table 2, zero-shot Qwen3-VL-8B (0.0118 SRCC) and Qwen3.5-35B (0.0880 SRCC) collapse to near-zero correlation, whereas Qwen2.5-VL-7B (0.3908) and Qwen3-VL-30B (0.3479) perform normally, suggesting output parsing failures rather than true capability differences.
      4. In Table 4 (cross-dataset evaluation), the middle block evaluates AesMind on PARA, but 50% of NeuroAes (12,500 images, including 10,000 training images in NeuroAes-Data) is directly sampled from PARA (Section 3.2). It is unclear whether the 10,000 PARA images in NeuroAes-Data were strictly excluded from the PARA test split evaluated in Table 4.
  3. Writing, formatting, and metadata issues:

    • The submission title on OpenReview has a typo at the end ("...Understanding and Assessmen", missing the final "t").
    • The number of human subjects in the subjective experiment is inconsistent across the paper: Figure 2 (page 4) states "Construct personal profiles of 71 subjects", whereas Appendix D.2 (page 20) states "We recruited a total of 50–60 participants".
    • In Table 12 (page 39), the "Peak Learning Rate" for Stage 1 (Skill Harnessing, which is inference-only prompt optimization without gradient updates) is listed as "4.0", and the image resolution is written as "256 x 322 to 384 x 322" (a formatting error for 256 x 32^2 to 384 x 32^2).
    • Several minor typos appear in the text and figures, such as "<imgae>" and "KL Divergency" in Figure 4, "NeuroS-Bench" in the captions of Tables 13 and 14, "chain-of-though" (page 3, line 161), "to generates" (page 4, line 214), and an incomplete sentence at the end of the Section 4 intro ("...and the subsections below", page 6, line 299).
Questions for authors
  1. Regarding the dataset annotations and Figure 9: Could the authors clarify how the human verification stage (Appendix D.3.2) was quality-controlled, and why the showcase example in Figure 9 contains stark contradictions between the Sensory-Motor tags ("cool tone", ["blue", "green", "gray"], "two" people, "daytime") and the actual sunset silhouette image and critique ("golden-orange tones", "lone silhouette at sunset")? What was the empirical modification/rejection rate by human annotators during verification?
  2. Regarding the Description task evaluation: Why do the LLM-as-a-Judge scores in Tables 3 and 6 exceed 1.0 (e.g., 1.1840 and 1.2328) when Eq. 7 in Appendix E.2.3 explicitly normalizes S_desc to [0, 1]? In addition, how do baseline models (e.g., Gemini-3-Pro, Qwen3.5-9B, and ArtiMuse) perform on NeuroAes-Bench (Perception and Description) when provided with the same task-specific Skill system prompts (especially the 50–80 word single-paragraph constraint for S_conc) or when ArtiMuse is fine-tuned on NeuroAes-Data as in Table 2?
  3. Regarding the Stage 3 GRPO trade-offs and reward design: In Table 6, why does adding GRPO (Row 7 vs. Row 6) reduce overall Perception accuracy by 5.17% (and why does Row 4 without Skill outperform Row 7 with Skill on both Perception and Description)? Does including multiple-choice and description samples in GRPO—or using LoRA instead of full-parameter fine-tuning during GRPO—mitigate this degradation? Furthermore, could the authors provide an ablation comparing the proposed hybrid regression reward (Eq. 3) and prior-adjusted categorical reward (Eq. 4) against standard L1/MSE and uniform accuracy rewards?
  4. Regarding the numerical discrepancies and experimental protocols:
    • Why are the Perception scores for the un-finetuned Qwen3.5-9B baseline different between Table 1 (42.79 / 19.86 / 42.05) and Table 6 Row 1 (53.21 / 19.28 / 45.65)?
    • In Table 4, for the main "AesMind (Ours)" row (without ⋆), was the model trained on NeuroAes-Data or on each benchmark's respective training set? If trained on NeuroAes-Data, did the authors verify zero overlap between the 12,500 PARA images in NeuroAes and the PARA test split?
11 comments

To Reviewer J49W:

Thank you for the constructive and detailed feedback. We appreciate your recognition of the human-rated VAD formulation and the multi-task evaluation, and address the concerns below.

1. W1: Methodology, annotation validity, and evaluation circularity

(a) & Q3.3: Methodological contribution and reward ablation

Thank you for pointing out that SkillOpt-based prompt optimization, LoRA SFT, and GRPO are established techniques, and that the original stage ablations did not distinguish the effects of our content rewards from those of standard RL post-training.

The technical contribution lies in how prompt-level skill harnessing interacts with parameter-level post-training, rather than in introducing these components individually. To our knowledge, their joint use for aesthetic understanding and assessment has not been explored previously. Importantly, their combination is not interchangeable: a skill useful before training need not benefit an already specialized model. Before finalizing AesMind, we explored alternative integration orders, including attaching the base-model skill after SFT or SFT + GRPO. We report these earlier results alongside the training-stage ablations in the table below and in Table 20 in the revised paper :

ModelSkill placementPerception (ACC)Assessment (SRCC)Assessment (PLCC)
Qwen3.5-9BNone45.650.25370.2613
Qwen3.5-9B + SkillInference only46.950.44920.4956
Qwen3.5-9B + SFTNone62.230.48710.5270
Qwen3.5-9B + SFT + SkillAttached after SFT70.530.44560.4941
Qwen3.5-9B + SFT + GRPONone66.210.65320.6707
Qwen3.5-9B + SFT + GRPO + SkillAttached after SFT + GRPO65.090.61690.6379
Qwen3.5-9B + Skill + SFTBefore SFT70.530.50960.5420
AesMind (Skill + SFT + GRPO)Before SFT (ours)65.360.69860.7261

The results show that skill placement matters most clearly for continuous assessment. Attaching the skill after specialization degrades assessment, even though it can improve perception after SFT; incorporating it before training improves assessment with both SFT and SFT + GRPO. During the same exploration, we also optimized new skills directly on the trained SFT and SFT + GRPO models. Neither search produced a candidate that passed the validation gate, which requires improvement over the current best on the full validation set. These findings show why simply stacking the components does not reproduce the benefits of our progressive design. Our methodological contribution concerns their integration for neuroaesthetic perception, continuous VAD assessment, and description, rather than novelty in the individual training primitives.

(a) & Q3.3: Methodological contribution and reward ablation (continued)

To validate the effectiveness of the proposed rewards beyond their design rationale, we have now added the standard-reward comparison requested in Q3. The proposed rewards are designed to address two specific challenges: distinguishing precise VAD predictions from merely close ones, and counteracting the dominance of frequent categories in photographic attributes.

We retrained Stage 3 by replacing the hybrid regression reward (Eq. 3) with an L1 reward and the prior-weighted categorical reward (Eq. 4) with uniform binary accuracy. The format reward, step-wise curriculum gate (Eqs. 5 and 8), training data, and hyperparameters were kept unchanged. This provides a comparison between standard and task-specific content rewards under the same GRPO recipe. Row numbers below follow Table 7 in revised paper (Table 6 in raw submission) , and "new" denotes the added run.

No.SkillLoRA SFTGRPO rewardMotor ACCEmotion ACCOverall ACCSRCCPLCCWord2Vec
(6)✓✓–76.3147.9570.530.50960.54200.9250
new✓✓L1 + uniform binary69.2337.7262.810.69060.70470.9254
(7)✓✓Ours (hybrid + prior-weighted)70.8943.7565.360.69860.72610.9262

As shown in the table above and in Table 22 in the revised paper, the proposed rewards outperform the simple alternatives on all reported metrics. Overall perception accuracy improves by 2.55 percentage points, with a 6.03-point improvement in Emotion accuracy; assessment SRCC/PLCC improve by 0.0080/0.0214. Since both variants use GRPO under the same training settings, these results support an empirical benefit from the task-specific reward design beyond the use of RL post-training alone. Skill + SFT still achieves the highest perception accuracy in this table, so the reward improvements mitigate, rather than eliminate.

Because one GRPO run takes at least five days on 8 × A100 GPUs, we replaced both rewards in a single run rather than ablating them separately. We omit the LLM-as-a-Judge Description metric here because its extensive API evaluation could not be completed within the available time and budget. After completing this evaluation, we will incorporate the full ablation results, the skill-placement experiments, and the corresponding analysis into the main text and supplementary material.

(b) & Q1: Annotation validity and human verification

Thank you for carefully examining Figure 9. We appreciate you bringing this issue to our attention. The contradictions stem from a version-control typographical error during the manuscript's figure preparation, where an intermediate, uncorrected annotation draft was mistakenly visualized instead of the final human-verified ground truth used for training.

We have thoroughly audited this sample and have provided the following clarifications regarding the exact progression of its labels:

Cool tone and blue/green/gray versus warm color impression: Our categorical annotation pipeline for color dimensions involves three stages: (1) Gemini initial draft, (2) Algorithmic cross-check, and (3) Human verification. For this image, Gemini initially predicted a "cool tone" with primary colors ["blue", "green"]. Our rule-based pixel-statistics algorithm then added "gray" to the primary colors while retaining the "cool tone". This intermediate Stage-2 result is what was inadvertently rendered in Figure 9. During the final Stage-3 human verification, human annotators correctly identified the discrepancy with the visual warm impression, changing the tone to a "mixed cool-warm" and adding "orange" to the primary color families. The actual NeuroAes training dataset contains this corrected, final version.

People count and gender: Upon zooming in on the original high-resolution image, two distinct figures can be clearly discerned, supporting the ground-truth count of "two". Similarly, the "male/female" gender tags shown in the figure reflect the uncorrected intermediate draft. Given the visual ambiguity of these silhouettes, gender cannot be reliably identified for both figures. During the final human verification, annotators corrected the tags to ["male", "uncertain"], as one silhouette can be roughly inferred as male while the other remains indistinguishable. The actual dataset correctly uses these verified tags.

Daytime versus sunset: In our taxonomy, the "daytime" category broadly covers daylight conditions, while "evening" primarily covers blue-hour conditions after sunset. While "daytime" is defensible under this rubric for a transitional sunset scene, we acknowledge the boundary ambiguity and have added clearer documentation for these definitions.

We have updated Figure 9 in the revised manuscript to reflect the correct, final human-verified annotations used in the actual dataset.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Our Responses & Paper Revisions

We sincerely thank all the reviewers for their valuable and constructive reviews. We appreciate that the reviewers acknowledged NeuroAes for providing a comprehensive neuroaesthetics-grounded framework (Reviewers 5B6G, D7zv) and a fine-grained, carefully constructed dataset and benchmark (Reviewers 5B6G, D7zv, J49W). They also recognized AesMind's comprehensive multi-task aesthetic understanding capabilities (Reviewers 5B6G, J49W), as well as its strong performance and cross-dataset generalization (Reviewers 5B6G, D7zv, J49W).

We have made our best effort to address the concerns and revise the paper accordingly. The major modifications are summarized as follows.

  1. Annotation Reliability: We have clarified the annotation sources and human verification procedures across the three tasks, providing additional details on participant backgrounds and field-specific guidelines. We also quantified human involvement, and corrected the corresponding annotation statistics and figure presentation.

  2. Methodological Analysis and Reward Ablation: We have added experiments to investigate the interaction between prompt-level Skill integration and parameter-level post-training, demonstrating the importance of skill placement. We further conducted a controlled reward ablation, providing empirical evidence for the effectiveness of our task-specific reward design over standard alternatives.

  3. Baseline and Generalization: We have incorporated additional comparisons with UniPercept and PortraitAes, together with Skill-assisted and in-domain fine-tuning baselines. We also evaluated the original AesMind checkpoint on VAB and AesBench without additional training or fine-tuning, providing further evidence of its generalization beyond NeuroAes.

  4. Analysis of Training-stage Trade-offs: We have expanded the analysis of Skill, SFT, and GRPO, clarifying their distinct and task-dependent effects on perception, assessment, and description. We further discussed the observed performance trade-offs and clarified the rationale for selecting the final assessment-oriented multi-task configuration.

  5. Numerical Consistency and Revisions: We have corrected numerical inconsistencies, clarified evaluation protocols and benchmark-specific settings, and addressed issues in VAD prediction validity, training configurations, benchmark descriptions, and manuscript presentation to improve the accuracy and reproducibility of our work.

The corresponding revisions and additional experimental results are highlighted in teal in the revised manuscript and supplementary material. Point-for-point responses to specific comments are given in the following reviewer-specific responses. We welcome any further discussions and will address any remaining concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.