acceptodds
Withdrawn from ICLR 2027

The Art of Retrieval Meets Learning: RL-Guided Best-of-N and Auxiliary-Head Fusion for Style-Transfer Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The rapid progress of generative imaging has made authenticity detection increasingly difficult, particularly when synthetic content also imitates established artistic styles. We introduce a reinforcement-guided Best-of- retrieval framework that recasts multimodal retrieval as sequential evidence acquisition. Rather than aggregating a fixed set of nearest neighbors, a lightweight policy learns when to retrieve additional evidence, switch modality, expand spatial context, reject unreliable candidates, abstain, or stop. Local visual representations are combined with LLM-generated surface descriptors, while the visual and textual encoders remain frozen at inference time. This formulation enables adaptive filtering of noisy or redundant retrieval evidence and improves style-aware representation separation. On AI-ArtBench, the proposed retrieval policy achieves 95.41% mean task accuracy across authenticity and style prediction, while also showing strong performance on additional real/synthetic and deepfake benchmarks. We further introduce an alternative trainable Auxiliary-Head Model that integrates global style, local artifacts, and cross-region coherence through a soft expert-fusion consistency regularizer, achieving 93.96% accuracy. Together, the two approaches provide distinct operating points for style-aware authenticity detection: adaptive retrieval with frozen encoders and trainable single-pass prediction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
style-transfer detection, AI-generated image detection, image authenticity, reinforcement learning, multimodal retrieval, Best-of-N retrieval, visual-textual fusion, deepfake detection, multimedia forensics

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_f9qtOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes two complementary methods for detecting AI-generated and style-transferred images. The first uses reinforcement learning to guide Best-of-N multimodal retrieval, combining local visual features with LLM-generated descriptions of artistic styles. An RL policy adaptively selects retrieval actions, adjusts spatial context, filters candidates, and determines when sufficient evidence has been collected. The second is a trainable fusion model with three branches targeting global style, local artifacts, and cross-region consistency. The authors also provide theoretical analyses and conduct comparative experiments, ablation studies, and robustness tests on AI-ArtBench, GenImage, and several deepfake datasets. The results suggest that adaptive retrieval improves detection performance, while the fusion model offers a single-pass alternative.

Critical strengths
  1. The RL-driven retrieval framework extends static multimodal retrieval into a sequential decision process, enabling adaptive retrieval, candidate filtering, and stopping. Combining local visual features with textual style descriptions also helps capture fine-grained artistic details.

  2. The paper evaluates the methods across AI-generated art, general synthetic images, and deepfake datasets, with ablation studies and analyses of cross-dataset generalization, efficiency, abstention, and calibration.

Critical weaknesses
  1. The RL-guided retrieval framework and auxiliary-head fusion model differ substantially in architecture, optimization, and inference, with no clear theoretical connection. The paper therefore feels like two separate studies rather than a cohesive work.

  2. The retrieval framework mainly applies standard PPO, while the fusion model relies on conventional multi-branch feature extraction. Moreover, the modest gains from RL and the lack of comparisons with simpler sequential selection methods leave its necessity and technical contribution unclear.

  3. Some derivations do not match the actual training objectives, particularly the consistency regularization analysis. The claims regarding fixed-point uniqueness, stability, and PAC-Bayes generalization also lack sufficient justification. Furthermore, improved class separation alone does not support the claimed reduction in visual encoder bias through multimodal fusion.

  4. The data used for VLM selection and the few-shot sampling rules are insufficiently described, raising concerns about potential information leakage. Several tables also show unusually regular numerical patterns, while the large gains from consistency regularization lack convincing explanations.

  5. Several baseline results are taken from prior studies without confirming consistent data splits, task definitions, or evaluation protocols. In addition, the reported inference times exclude online text generation costs, making the practical efficiency of the method difficult to assess.

Questions for authors
  1. Appendix A.7 analyzes a loss function different from the one used in the proposed method, and the uniqueness of the consensus fixed point remains unproven. How do these results apply to the actual model, and do the convergence and generalization claims still hold?

  2. Table 3 shows unusually regular numerical patterns across encoders, preprocessing methods, and retrieval settings, particularly the nearly constant gaps between variants with and without text. Could the authors explain these patterns and provide raw logs or per-sample predictions?

  3. The captioner is selected using label-conditioned embedding separability, while few-shot examples come from the retrieval bank. Please clarify the data splits, few-shot selection rules, and whether query samples were excluded from retrieval banks during policy training.

  4. The RL policy achieves only modest gains over the non-RL MLP aggregator. Are these gains due to sequential optimization or simply additional evidence and computation? Comparisons with simpler learned sequential selection methods would help justify the use of RL.

  5. Removing the consistency loss causes a substantial performance drop. What explains such a large effect from a simple regularizer? Were all other architectural settings, hyperparameters, and training procedures kept identical?

  6. Several baseline results are taken from prior studies without verifying consistent protocols. Do these methods use the same data splits, style taxonomies, task definitions, and evaluation settings? If not, which performance comparisons remain valid?

Reviewer_5pEaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes two independent paths for identifying artistic images. The first step is to use local visual features and surface texture descriptions generated by VLM for retrieval. The PPO strategy is used to decide whether to continue retrieval, switch modes, expand regions, reject candidates, reject answers, or stop, and ultimately weight the retained evidence through voting. The second one is a trainable multi head classifier that integrates global style, local artifacts, and regional consistency, and increases the consistency loss between fusion prediction and expert average prediction. The author reported that the average accuracy of the search method for authenticity and style on AI ArtBench was 95.41%, and the accuracy of the auxiliary head model was 93.96%.

Critical strengths
  1. Adaptive evidence acquisition has a reasonable motivation. The authenticity and style judgment of artistic images may require evidence from different regions and modalities; Learning when to add evidence, discard candidates, and stop is more flexible than fixed aggregation of all neighbors.

  2. Considered selective prediction and retrieval behavior. The analysis of action utilization, action ablation, retrieval library size, and quality before and after filtering also provides information for understanding strategic behavior.

Critical weaknesses
  1. There is no unified task and indicator for core performance comparison. The 95.41% accuracy of RL is the average of the true false accuracy rate of 97.41% and the style accuracy rate of 93.41%; The auxiliary head model in Section 3.2 only defines binary classification outputs and BCE, without specifying how to generate corresponding style classification results. Therefore, Table 2 places 93.96% of the auxiliary head and 95.41% of the RL in the same Accuracy row, and the comparative meaning is unclear

  2. There are numerical errors in the main table that can be directly verified, and some experimental explanations are also contradictory. In Table 3, the Mean Task Accuracy in the following two rows is inconsistent with the corresponding task accuracy: k=5, 44 and k=9, 55

  3. The consistency theory analysis uses different losses, and the contraction mapping conclusion is incorrect.

Questions for authors
  1. What labels and indicators are predicted by each method in Table 2? Can you provide the re evaluation results of the same agreement and the original statistics of Table 3?

  2. Which type of consistency loss is used in actual training? How to correct the unique fixed point, stability, and calibration guarantee?

  3. What work level deduplication has AI ArtBench and WikiArt done? What is the generalization result after deduplication?

  4. Can you provide a complete PPO environment definition and end-to-end cost including captioner?

Reviewer_2WUMOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper combines adaptive retrieval-based inference with an auxiliary-head detector for style-aware authenticity assessment. The retrieval component uses frozen visual and textual encoders and a reinforcement-learning policy that selects, expands, rejects, or stops collecting evidence. The detection component combines global-style, local-artifact, and cross-region-coherence predictions through learned fusion and consistency regularization. Validation includes artistic and general synthetic-image benchmarks, deepfake datasets, non-RL retrieval controls, action and reward ablations, selective prediction, cross-dataset transfer, and theoretical analysis.

Critical strengths

The controlled retrieval comparisons provide some genuine support for adaptive evidence selection. The sequential policy consistently outperforms fixed voting, learned non-sequential fusion, and greedy confidence-based selection, with paired statistical analysis reported for the principal comparisons. Cross-dataset results also favor the adaptive policy over the learned non-sequential alternative. These findings suggest a modest practical benefit beyond merely adding retrieval or model capacity.

Critical weaknesses

Several theoretical claims are unsound or do not analyze the implemented algorithm. The majority-vote argument assumes independent evidence despite correlated neighbors and adaptive selection, while the margin analysis largely restates a property of correct nearest-class prediction. More seriously, the consistency appendix analyzes an inter-head variance objective that differs from the implemented fusion-to-expert-mean loss, so its convergence and generalization conclusions do not follow. The PAC-Bayes analysis also leaves query-dependent retrieval and data-dependent prior concerns unresolved. These issues invalidate the claimed guarantees rather than merely limiting their generality.

The evaluation does not adequately support the broad detection and generalization claims. Split construction and group disjointness are insufficiently specified, the referenced leave-one-style-out results are absent, and the main cross-dataset table does not clearly distinguish per-dataset training from transfer. Captioner selection using ground-truth-label-conditioned distances also raises a potential test-label leakage concern unless the procedure is strictly restricted to training data.

Questions for authors

What are the exact dataset splits; are related frames, artists, generators, and source images group-disjoint; and where are the referenced leave-one-style-out results?

Were the reported detectors trained independently within each dataset or transferred from the artistic benchmark, and can matched external baselines be evaluated under identical protocols?

Are captioner selection and all few-shot caption exemplars strictly restricted to training images and training labels?

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

Related papers

Discussion (0)

Sign in to comment.