acceptodds
Under review as a conference paper at ICLR 2027

Latent Aesthetic Deliberation: Learning to Refine Aesthetic Judgments

Abstract

Image Aesthetic Assessment (IAA) requires weighing visual cues whose aesthetic contributions depend on image context. Although image–rating pairs are widely available, they specify judgment outcomes rather than how evidence should be integrated or assessments revised. Direct scoring with multimodal large language models (MLLMs) does not explicitly model judgment revision, whereas supervised verbalized reasoning requires additional reasoning-text targets and incurs token-by-token decoding overhead. Motivated by this, we introduce Latent Aesthetic Deliberation (LAD), a recurrent framework for learning judgment refinement from ratings without annotated reasoning trajectories. After a single multimodal encoding pass, an evolving judgment state repeatedly retrieves and integrates evidence from persistent image-token memory. Regression against human ratings trains the recurrent scorer, while internal self-distillation aligns intermediate score readouts with the final prediction without prescribing their latent representations. At inference, LAD fuses the recurrent score with a parallel language-aligned rating estimate without decoding an explanation. Extensive experiments demonstrate state-of-the-art performance on the evaluated benchmarks for holistic and aspect-specific aesthetic assessment.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multimodal Large Language Model, Implicit Chain-of-thought, Interpretable Visual Aesthetic Evaluation

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_De7YOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Latent Aesthetic Deliberation (LAD), a recurrent framework for image aesthetic assessment that refines aesthetic judgments in latent space without requiring explicit chain-of-thought annotations or textual reasoning during inference. After a single multimodal encoding pass, LAD maintains persistent image-token memory and iteratively updates a compact latent judgment state by repeatedly retrieving and integrating visual evidence. The model is trained using human mean opinion scores together with internal self-distillation to supervise intermediate score predictions, and the final aesthetic score combines the recurrent regression output with a language-aligned rating estimate. Experiments on AVA, PARA, AADB, and ICAA17K show strong performance for both holistic and aspect-specific aesthetic assessment, with ablations supporting the effectiveness of recurrent visual evidence retrieval and intermediate trajectory supervision.

Critical strengths
  1. LAD introduces a well-motivated latent deliberation mechanism that iteratively refines aesthetic judgments without requiring explicit CoT annotations or costly textual reasoning at inference.

  2. The combination of persistent visual memory, recurrent evidence retrieval, and intermediate self-distillation is technically coherent and allows the model to learn multi-step judgment refinement using only final human MOS supervision.

Critical weaknesses
  1. The paper repeatedly frames the recurrent process as deliberation, reasoning, and judgment refinement, but the supervision does not actually establish a reasoning process. Intermediate steps are only encouraged to match the final prediction, which itself is supervised by the MOS, so the method may simply perform iterative feature refinement or repeated score calibration rather than meaningful aesthetic reasoning. The current evidence does not convincingly distinguish these interpretations.
Questions for authors
  1. Could the authors evaluate LAD on more diverse or out-of-domain aesthetic benchmarks, such as VAB [1] and AesBench [2]? In particular, VAB is grounded in expert consensus, so I would be interested in whether the gains on conventional MOS-based benchmarks transfer to expert-aligned aesthetic judgment.

  2. The final score is computed as 0.3scls+0.7sreg{0.3s_{cls} + 0.7s_{reg}}. Is there a clear ablation comparing the classification branch alone, the recurrent regression branch alone, and their fusion, as well as sensitivity to the fusion weight γ{\gamma}? Otherwise, it is difficult to tell how much of the gain comes from LAD itself versus the score fusion, and why γ=0.3{\gamma=0.3} is the appropriate choice.

  3. The paper describes the recurrent process as deliberation and judgment refinement, but the intermediate states are only supervised to match the final prediction. Is there additional evidence that these steps capture meaningful aesthetic reasoning rather than simply iterative feature refinement or score calibration?

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

3 comments
AuthorsOpenReview ↗

Response to Reviewer De7Y

Response Summary: We thank the reviewer for recognizing LAD's motivation and coherent design. The constructive feedback has helped us improve the presentation and strengthen the empirical evidence. We address the concerns with additional mechanism interventions, zero-shot VAB evaluation, and fusion analyses, while clarifying the evidence supporting aesthetic deliberation.

Q1: Evaluation beyond MOS-based benchmarks

We evaluated LAD on VAB without additional training, following its individual-image scoring protocol: average three scores per image, then select the highest- and lowest-scoring images within each candidate set.

MethodTop-1 (%)Bot-1 (%)TB-1 (%)
Q-Align52.247.838.2
ArtiMuse44.251.234.8
Q-Instruct38.551.531.2
LAD52.848.238.8

LAD improves over Q-Align by 1.15%, 0.84%, and 1.57%, respectively, and leads Top-1/TB-1 among these scorers, providing evidence of transfer to expert comparative judgments. Our AesBench access request remains pending. If access arrives in time, we will evaluate LAD and update the results. Both benchmarks are discussed in the revised manuscript.

Q2: Branch contributions and fusion weight

We compared regression-only scoring with two fusion weights on VAB, using s^=γscls+(1−γ)sreg\hat{s}=\gamma s_{\mathrm{cls}}+(1-\gamma)s_{\mathrm{reg}} (Supplementary PDF E.3.1, Table 10).

ClassificationRegressionTop-1 (%)Bot-1 (%)TB-1 (%)
0.01.051.748.337.8
0.30.752.848.238.8
0.50.553.047.839.2
1.00.053.547.539.8

Default fusion improves Top-1/TB-1 over regression-only scoring by 2.13%/2.65%, while Bot-1 decreases by 0.1 percentage point. Compared with classification-only scoring, it improves Bot-1 by 0.7 percentage points, while Top-1/TB-1 decrease by 0.7/1.0 percentage points. These results reveal a trade-off between best- and worst-image selection across the two branches and their fusion.

Why γ=0.3\gamma=0.3. The rating branch converts aesthetic-level probabilities into a language-aligned score. The recurrent branch learns fine-grained judgment updates through visual evidence integration and direct MOS supervision. We set γ=0.3\gamma=0.3 across datasets with no purposeful search. We aim to prioritize the recurrent prediction while retaining rating information.

W1 & Q3: Evidence for aesthetic deliberation

Score consistency alone does not establish aesthetic reasoning. Final MOS regression backpropagates through the complete recurrence. Self-distillation aligns intermediate scalar readouts with the detached final prediction, without requiring latent representations to match. The evolving state carries information into subsequent visual queries. Scalar scores are readouts, not feedback inputs.

We tested intermediate-state interaction and continued visual access through two inference-time interventions on PARA, preserving the normally computed first-step state and intervening from step 2 onward.

ConfigurationSRCCPLCC
Full LAD0.9410.958
Block score-token attention to thought tokens0.9330.949
Zero subsequent visual cross-attention updates0.9250.939

The first blocks the score token from attending to thought tokens during self-attention, retaining its own state and direct visual-memory access. The decrease supports the contribution of thought-to-score information transfer.

The second zeros subsequent visual cross-attention updates for all tokens, retaining initialization and first-step information as well as later self-attention, feed-forward, and gated updates. Accuracy decreases despite continued latent processing, supporting the contribution of renewed visual access. Details are in Supplementary Appendix H.5, Table 18.

These results complement Figures 3 and 4: Figure 3 shows regression-score revisions toward MOS in selected examples. Figure 4 depicts changing visual-memory attention across steps. The figures describe how predictions and retrieval evolve. The interventions show that disrupting these information pathways reduces the trained model's accuracy. Full-test-set trajectories also reflect dataset-level statistics on reasoning process (Supplementary Appendix G.3, Table 14). Together, these findings support learned, visually grounded aesthetic judgment refinement through state interaction and continued evidence integration.

Closing: We thank the reviewer again for helping us clarify LAD's mechanism and strengthen its supporting evidence.

Thank you for the detailed response and the additional experiments. Although some of the improvements are relatively modest, I believe the new evidence sufficiently supports the overall claims of the paper and addresses my main concerns. I appreciate the authors' efforts and will raise my score accordingly.

AuthorsOpenReview ↗

Thank You for Your Reply

Thank you for carefully considering our response and additional experiments. We appreciate your positive reassessment and are glad that the new evidence addresses your main concerns. Your constructive feedback has helped us strengthen the paper, and we sincerely appreciate the time and care you have devoted to reviewing our work.

Reviewer_Q885OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Latent Aesthetic Deliberation (LAD) for image aesthetic assessment. Instead of directly predicting a score or generating explicit Chain-of-Thought, LAD maintains a recurrent latent judgment state that repeatedly retrieves evidence from persistent image tokens and progressively refines the aesthetic score. Intermediate score estimates are regularized through self-distillation, and the final recurrent regression score is fused with a language-aligned rating estimate. The method achieves strong results on both holistic and aspect-specific aesthetic benchmarks.

Critical strengths
  1. The method has a clearer technical contribution than simply combining standard MLLM training techniques: the persistent visual memory, evolving latent judgment state, recurrent evidence retrieval, and intermediate score supervision form a coherent design.
  2. The paper is well motivated: LAD avoids both explicit reasoning annotations and token-by-token CoT decoding.
  3. The empirical results are strong and consistent. LAD substantially improves over prior methods on AVA and AADB, while also achieving the best results on composition, lighting, and color assessment.
Critical weaknesses
  1. The notion of “latent deliberation” is stronger than what is empirically demonstrated. Recurrent cross-attention and hidden-state refinement show iterative computation, but do not necessarily establish that the model performs meaningful multi-step aesthetic reasoning.
  2. The trajectory analysis relies heavily on a small number of selected qualitative examples. A dataset-level analysis showing how often additional iterations actually improve prediction would provide stronger evidence for progressive refinement.
  3. The final prediction uses a fixed fusion weight γ\gamma between the recurrent regression score and the classification score; a more fine-grained or adaptive fusion strategy may be worth exploring.
Questions for authors
  1. Can the authors provide a dataset-level statistic showing how frequently each recurrent step improves the prediction over the preceding step, rather than only selected trajectory examples?
  2. Since computational efficiency is a key motivation, how does LAD compare with one-shot scoring and verbalized-CoT methods in inference latency or FLOPs?
  3. How sensitive is performance to the fusion weight γ\gamma, and how much of the final gain comes from latent deliberation versus regression–classification score fusion?
2 comments
AuthorsOpenReview ↗

Response to Reviewer Q885 - Part I (W1,W2,Q1)

Response Summary: We thank the reviewer for recognizing LAD’s coherent design, motivation, and strong results across holistic and aspect-specific aesthetic assessment. Your constructive feedback helped us clarify latent aesthetic deliberation and strengthen its supporting evidence. We address the concerns through mechanism ablations, dataset-level trajectory statistics, computational-cost comparisons, and fusion analyses, clarifying their implications for visual evidence integration and aesthetic judgment refinement.

W1: Evidence supporting latent aesthetic deliberation

Thank you for highlighting the distinction between iterative computation and meaningful aesthetic reasoning. Recurrence alone does not establish the latter. In LAD, aesthetic deliberation refers to a learned process in which an evolving judgment guides renewed access to visual evidence, and the retrieved evidence informs subsequent assessment. This formulation addresses the context-dependent nature of aesthetics: a visual cue must be evaluated in relation to other cues and the overall image. To examine whether these interactions contribute to prediction, we conducted two inference-time interventions on PARA using the same trained model and the same normally computed first-step state. Both interventions begin at the second recurrent step, without retraining.

ConfigurationSRCCPLCC
Full LAD0.9410.958
Block score-token attention to thought tokens from step 20.9330.949
Remove subsequent visual cross-attention updates from step 20.9250.939

Contribution of intermediate-state interaction. The first intervention prevents the score token from attending to the thought tokens during subsequent self-attention updates. It retains its own evolving state and direct access to visual memory. SRCC and PLCC decrease by 0.85% and 0.94%, respectively, relative to full LAD. Thus, direct visual access by the score token does not preserve the full model’s accuracy when information transfer from the thought tokens is blocked. This supports the functional contribution of the additional evolving states to aesthetic assessment.

Contribution of continued visual evidence access. The second intervention sets the cross-attention visual updates of all latent tokens to zero from step 2 onward. It preserves visual information obtained during initialization and the first step, together with subsequent self-attention, feed-forward, and gated updates. SRCC and PLCC decrease by 1.70% and 1.98%, respectively. Continued processing of the accumulated states therefore does not preserve the full model’s accuracy when subsequent visual updates are removed.

These interventions complement the analyses in main PDF Sections 4.3.1–4.3.2. Figure 3 illustrates score revisions toward human MOS across recurrent steps in selected examples, while Figure 4 visualizes step-dependent changes in attention to persistent visual memory. These qualitative observations describe how predictions and evidence access evolve; the interventions further show that blocking score-token access to thought tokens or removing subsequent visual updates reduces the trained model’s accuracy. Together, these findings support LAD’s learned, visually grounded aesthetic judgment refinement through intermediate-state interaction and continued evidence integration. Detailed performance and parameter comparisons are reported in Supplementary Section H.5, Table 18.

W2 & Q1: Dataset-level analysis of step-wise prediction improvement

Thank you for requesting a dataset-level analysis beyond the selected qualitative examples. We extended the trajectory analysis to all 3,000 images in the PARA test set, measuring how frequently each successive prediction moves closer to the ground-truth MOS.

The following statistics concern the fused predictions. For each image, we compare the absolute prediction error between successive steps and classify the change as improved, degraded, or tied within the evaluation tolerance.

ComparisonImprovedDegradedTied
Step 2 versus step 169.07%16.53%14.40%
Step 3 versus step 260.20%25.43%14.37%
Step 4 versus step 350.27%34.33%15.40%
Final step versus first step63.73%29.07%7.20%

Improvements outnumber degradations at every transition, with improvement rates of 69.07%, 60.20%, and 50.27%. From the first to the final step, absolute error decreases for 63.73% of images, increases for 29.07%, and remains tied for 7.20%. Across complete trajectories, 45.70% improve at every transition, 60.10% never deteriorate, and 75.10% improve at least once. These full-test-set results support recurrent aesthetic judgment refinement beyond Figure 3’s selected examples, without implying monotonic improvement for every image. Detailes are reported in Supplementary Sec. G.2, Table 13; G.3 Table 14; G.4, Table 15.

AuthorsOpenReview ↗

Response to Reviewer Q885 - Part II (Q2,Q3,W3)

Q2: Computational cost relative to one-shot scoring and verbalized CoT

Thank you for requesting a quantitative efficiency comparison. We measured the computational cost of Q-Align and LAD and provide published profiles of verbalized CoT image-quality assessment models as additional references.

Here, TFLOPs denotes the total floating-point operations required to complete one inference, measured in trillions (10¹² operations). It measures computational workload: when predictive performance is comparable, lower TFLOPs indicate a lower computational cost per inference.

MethodParametersInput sizeTFLOPsLatency (s)
Q-Insight (CoT)8.29B512 × 3848.7102.720
VisualQuality-R1 (CoT)8.29B512 × 3847.7402.340
Q-Align8.198B448 × 4482.0620.162
LAD (T = 4)8.206B448 × 4482.0670.164

The Q-Insight and VisualQuality-R1 values are literature-reported references; they use a different input resolution and are not a controlled, same-environment comparison with LAD.

Relative to Q-Align at the same input resolution, LAD adds 0.24% FLOPs and 1.23% latency, corresponding to approximately 2 ms per prediction in our measurements. LAD evaluates the multimodal backbone once, followed by four updates of a compact latent state, without autoregressive explanation decoding. This allows repeated visual evidence integration with a small measured computational overhead over the one-shot scorer. Detailed performance and parameter comparisons are reported in Supplementary Section F.3, Table 12.

W3 & Q3: Fusion-weight sensitivity and the contribution of score fusion

Thank you for raising these related questions. We evaluated regression-only scoring and different fusion weights on the Visual Aesthetic Benchmark (VAB). Following the manuscript’s notation, the final score is s^=γscls+(1−γ)sreg\hat{s}=\gamma s_{\mathrm{cls}}+(1-\gamma)s_{\mathrm{reg}}, where γ\gamma is the classification-branch weight.

Classification weightRegression weightTop-1 (%)Bot-1 (%)TB-1 (%)
0.01.051.748.337.8
0.3 - default0.752.848.238.8
0.50.553.047.839.2

Top-1 and Bot-1 measure correct identification of the best and worst images within a group; TB-1 requires both selections to be correct.

Contribution of fusion. Relative to the recurrent regression branch alone, the default fusion improves Top-1 and TB-1 by 2.13% and 2.65%, respectively, while Bot-1 decreases by 0.1 percentage point. Fusion therefore provides an incremental benefit for identifying the best image and jointly identifying both extremes, with a small trade-off in worst-image selection. This comparison quantifies the additional contribution of output fusion on VAB; the recurrence ablations and visual-access interventions separately examine the recurrent scoring mechanism.

Sensitivity to the fusion weight. Increasing γ\gamma from 0.3 to 0.5 changes Top-1, Bot-1, and TB-1 by +0.2, −0.4, and +0.4 percentage points, respectively. The differences between these tested fusion weights are limited, although the preferred weight depends on the evaluation metric. We retain the original fixed coefficient as a consistent scoring rule across datasets. An adaptive fusion strategy is a worthwhile extension, while the present results characterize the behavior of fixed fusion. Detailed performance and parameter comparisons are reported in Supplementary Section E.3.1, Table 10.

Closing: We thank the reviewer for these suggestions, which have strengthened the fairness, repeatability, and component-level evaluation of our approach. Thank you again for your time, effort, and constructive feedback; we welcome any follow-up questions and will respond promptly.

Reviewer_9DuGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Latent Aesthetic Deliberation (LAD), a recurrent latent scoring approach for image aesthetic assessment using a multimodal large language model (MLLM). A single MLLM pass produces contextualized visual tokens, a task-conditioned score context, and rating logits. LAD initializes four thought tokens and one score token, then repeatedly updates them using a shared block with self-attention, cross-attention to persistent visual memory, and a gated feed-forward update. The final continuous score is learned from human mean opinion scores (MOS), while trajectory self-distillation encourages intermediate score readouts to agree with the final prediction. At inference, the recurrent score is fused with a rating-classification estimate without generating an explicit reasoning explanation. Experiments on AVA, PARA, AADB, and ICAA17K cover holistic and attribute-specific aesthetic prediction, with reported improvements over the evaluated baselines and additional component and recurrence-depth ablations.

Critical strengths

Strength 1: Strong empirical performance across aesthetic assessment benchmarks LAD reports substantial gains on established image aesthetic assessment benchmarks. Table 1 shows SRCC/PLCC of 0.927/0.933 on AVA and 0.901/0.887 on AADB, compared with 0.833/0.832 and 0.783/0.789 for ROC4MLLM, respectively. The results are also competitive on PARA and cover attribute-specific composition, lighting, and color assessment. These consistent benchmark results provide meaningful evidence for the predictive utility of the complete LAD framework.

Strength 2: A coherent, rationale-free scoring pipeline with controlled component evidence LAD obtains a continuous aesthetic score using one MLLM encoding pass followed by a compact shared recurrent module, without autoregressively decoding intermediate CoT text for scoring. The supplement specifies the implementation, loss routing, and component ablations. Across AVA, PARA, and AADB, removing the gate or trajectory term and replacing token-level memory with mean pooling reduces reported correlation performance. These observations support the practical coherence and effectiveness of the implemented pipeline, even though they do not establish that each specific design is necessary relative to other reasonable architectures.

Critical weaknesses

Weakness 1: Necessity of the multi-token latent state is not established LAD uses four thought tokens plus one score token, but the paper does not report an ablation of latent-token count or a comparison with a single-score-query cross-attention scorer. The mean-pooling ablation changes the visual memory rather than the number of latent queries. Consequently, the added value of the four thought tokens over a simpler query-based scoring head remains unclear. This matters because the separate evolving latent state is central to the proposed architectural motivation.

Weakness 2: The particular trajectory self-distillation target is insufficiently justified The no-trajectory ablation demonstrates that intermediate score consistency helps relative to removing the trajectory term. However, the training set already provides ground-truth MOS, and the paper does not compare the detached-final-score teacher with direct MOS supervision at intermediate readouts (deep supervision). Thus, the experiment supports the effectiveness of an auxiliary trajectory loss but not the necessity or relative advantage of the proposed self-distillation target, a stated methodological contribution.

Weakness 3: Shared recurrent computation is not isolated from additional network depth The T=1–5 experiment on PARA shows that models trained with more recurrent updates generally achieve better scores through T=4, but it does not compare four shared recurrent updates with a four-layer attention scorer using independent parameters at matched depth and visual input. Therefore, the results establish a benefit from greater computation within LAD, not a distinct advantage of weight sharing over conventional deep attention-based refinement. A fair comparison should report both performance and parameter/computation costs.

Questions for authors

Question 1: Latent-token necessity Could the authors compare the present five-token state with a single score query and several smaller latent-token configurations, keeping the backbone, visual memory, training data, and scoring protocol consistent? What measurable benefit do the four extra thought tokens provide?

Question 2: Self-distillation versus direct intermediate supervision Could the authors compare the current detached-final-score trajectory objective with direct ground-truth MOS supervision for the same intermediate readouts, matching loss type and weights as closely as possible? Does self-distillation offer an advantage beyond adding intermediate supervision?

Question 3: Shared recurrence versus independent layers Could the authors compare T=4 shared recurrent updates with an equally deep, non-weight-shared attention scorer that accesses the same visual memory? Please report SRCC/PLCC together with parameter counts and inference cost to distinguish gains from depth and gains from weight sharing.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 9DuG

Response Summary: We thank the reviewer for the constructive feedback and recognition of LAD's clarity, strong empirical performance, and coherent scoring pipeline. The constructive feedback has helped us improve the quality of our article. We added controlled experiments to clarify the benefits of the multi-token state, self-distillation, and shared recurrence.

W1 & Q1: Benefit of the multi-token latent state

We conducted two comparisons on PARA, keeping the backbone, token-level visual memory, training data, training configuration, and scoring protocol unchanged. Every configuration retains one score token. Comparing zero-thought-token scorers at T=1T=1 and T=4T=4 tests repeated updates. Varying thought-token count at fixed T=4T=4 tests the additional latent states.

ConfigurationThought tokensStepsSRCCPLCC
Single-step cross-attention scorer010.9200.932
Single-score-token recurrent scorer040.9290.942
LAD variant140.9350.951
LAD variant240.9380.956
LAD variant340.9400.958
LAD (default)440.9410.958
LAD variant640.9420.959
LAD variant840.9430.959

At fixed T=4T=4, four thought tokens improve SRCC/PLCC by 0.012/0.016 over the single-score-token scorer. This baseline already retrieves from the same visual memory recurrently, so the comparison isolates the added benefit of the multi-token state. Unlike mean pooling, it preserves token-level memory.

The first two rows also show a benefit from repeated updates with the latent-state size unchanged. Together, these comparisons support additional thought tokens beyond recurrence alone. Increasing thought tokens from four to eight adds 0.002/0.001, making four a compact default near the observed performance plateau. Details are in Supplementary Section H.7, Table 20.

W2 & Q2: Self-distillation versus direct intermediate MOS supervision

We added a deep-supervision baseline that replaces sg(sT)\mathrm{sg}(s_T) with ground-truth MOS at the same intermediate readouts. Only the supervision source changes: other settings remain unchanged.

Intermediate supervisionSRCCPLCC
Direct ground-truth MOS supervision0.9390.954
Detached final-score self-distillation (LAD)0.9410.958

Self-distillation improves SRCC/PLCC by 0.21%/0.42%, respectively. These modest gains support the chosen target beyond simply adding an intermediate loss.

Final MOS regression anchors the final judgment to human ratings and backpropagates through the entire recurrence. Self-distillation uses that detached final prediction, informed by the full sequence of evidence integration, as the reference for earlier readouts. It encourages consistency across recurrent assessments without annotated judgment trajectories. Details are in Supplementary Section H.6, Table 19.

W3 & Q3: Shared recurrence versus independent layers

We compared shared recurrence with four independent blocks on PARA. Both execute four structurally identical updates and access the same persistent visual memory at every step. Only parameter sharing changes. The backbone, training data, and remaining training and evaluation settings are unchanged.

ConfigurationTrainable parametersSRCCPLCC
Four independent blocks86.35M0.9380.953
Shared recurrent block (LAD)75.32M0.9410.958

At matched depth, LAD improves SRCC/PLCC by 0.32%/0.52% while using 12.8% fewer trainable parameters. This supports a shared judgment-update rule beyond the benefit of additional depth. Performance and trainable-parameter results are in Supplementary Section H.8, Table 21.

We also measured inference cost at 448×448448\times448 input resolution:

ConfigurationTotal parametersTFLOPs per inferenceLatency (s)
Four independent blocks~8.217B2.0670.165
Shared recurrent block (LAD)8.206B2.0670.164

TFLOPs denotes trillions of floating-point operations per prediction. Both models execute the same number of operations with matching tensor shapes, so sharing reduces stored parameters while reported FLOPs remain identical. Latency differs by 1 ms (0.61%), indicating comparable inference time. Thus, shared recurrence yields higher correlations with fewer parameters at comparable inference cost.

Closing: We thank the reviewer again for helping us strengthen the evidence for LAD's design choices.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Summary of Contributions and Additional Evidence

We thank the reviewers for their careful reading and constructive feedback.

Aesthetic scoring requires weighing visual cues in the context of the whole image, yet current labeling in existing datasets leaves open how those cues should be combined or revisited. We therefore ask whether a model can learn to revise its judgment from available ratings without requiring annotated reasoning steps during training or generating an explanation at inference. Our contribution is to couple visual evidence selection with continuous score learning in a single recurrent judgment process. LAD encodes the image once, and each evolving judgment state both guides the next read of the image tokens and yields a score through a shared readout. This places successive judgments on a common scale while leaving each state free to integrate different visual cues, giving LAD a common framework for both holistic and attribute-specific aesthetic assessment.

The reviewers found this goal well motivated and highlighted the link between evidence retrieval and score supervision as a coherent technical contribution, supported by controlled ablations, substantial gains in overall scoring on AVA and AADB, and strong results for composition, lighting, and color.

In the rebuttal, we provide additional comparisons and analyses to clarify which design choices contribute to LAD’s performance, how often recurrent updates improve a judgment, and what they cost at inference.

  1. Controlled comparisons of the latent state and its supervision. Varying the number of thought tokens at fixed depth shows that a multi-token state improves scoring over a single recurrent score query. A separate comparison with direct human-score supervision at every step shows a modest benefit from using the final recurrent prediction to guide earlier estimates (Supplementary Tables 19–20).

  2. Evidence for aesthetic judgment refinement. Interventions in the trained model show that later predictions benefit from both revisiting the image tokens and exchanging information within the latent state. Across the full PARA test set, revisions to the fused score moved predictions closer to human ratings more often than they moved them farther away. Together, these findings support our interpretation of latent deliberation as evidence-guided score revision (Supplementary Tables 13–14 and 18).

  3. Weight sharing and inference cost. At matched depth, shared updates achieved higher correlations than independent layers with fewer trainable parameters and comparable inference cost. We also measured a small overhead in FLOPs and latency over one-shot Q-Align and provided published verbalized-CoT measurements from other settings for context (Supplementary Tables 12 and 21).

  4. Branch contributions and transfer to expert judgments. Our zero-shot VAB evaluation extends the comparison to expert judgments, where testing each branch alone and at different fusion weights revealed a trade-off between best- and worst-image selection. We clarified that the fixed default was chosen heuristically to prioritize recurrent scoring. Access to the other suggested benchmark, AesBench, was pending at the time of our response.

We truly thank the reviewers and the Area Chair once again for the time and effort they have devoted to evaluating our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.