acceptodds
Withdrawn from ICLR 2027

Self-Evolving Experts for Continual Test-Time Adaptation in AI-Generated Image Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

AI-generated image detectors are typically trained once and kept static after deployment, making them vulnerable to emerging generators with unseen forensic traces. While continual test-time adaptation (CTTA) is a promising direction, effective self-training requires reliable pseudo-supervision and stable knowledge consolidation across distributions. We propose SEED (Self-Evolving Experts for Detection), a framework that enables pretrained detectors to adapt continually via a dynamically managed pool of residual experts, in the absence of ground-truth labels, generator identities, or domain boundaries. To this end, SEED dynamically manages its expert pool through shift detection for expert creation and decision-orthogonal expert routing. It creates a new expert only when the detector's scores show a persistent change, while using decision-orthogonal features to reuse existing experts when a context recurs. Each expert accumulates feature statistics weighted by the reliability of its pseudo-labels, and updates its residual head through balanced feature replay. Across four public benchmarks, SEED outperforms SOTA methods and improves final accuracy by up to 17.42 percentage points over the frozen detector without CTTA. All code and data will be publicly available upon acceptance.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
AI-generated image detection, continual test-time adaptation, continual learning

Reviews

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

This paper proposes SEED, a continual test-time adaptation framework for AI-generated image detection without access to target labels, generator identities, or explicit domain boundaries. SEED freezes the base detector and maintains a dynamic pool of residual experts. Dual-Space Expert Modeling (DSEM) combines base-logit shift detection for expert creation with decision-orthogonal features for expert routing. Authenticity Uncertainty Estimation (AUE) derives pseudo-labels and reliability weights from Gaussian mixture models of base logits, while Progressive Knowledge Consolidation (PKC) maintains reliability-weighted feature statistics for balanced feature replay. Experiments on four benchmarks report improvements in online and final holdout performance, including up to a 17.42-percentage-point gain in final accuracy over the frozen detector. Additional ablations examine the contributions of the proposed components and expert reuse.

Critical strengths
  1. The paper addresses a practically relevant continual adaptation problem. AI-generated image detectors must accommodate evolving generators while retaining previously acquired detection capabilities. Studying this problem without target labels or explicit domain boundaries is valuable for deployment. The prediction-before-update protocol and holdouts that are disjoint from the complete adaptation stream strengthen the evaluation.

  2. The framework provides a coherent integration of adaptation and retention mechanisms, supported by consistent empirical gains. Dynamic residual experts, reliability-weighted pseudo-supervision, and statistical feature replay offer a plausible approach to adapting a frozen detector without retaining historical images. SEED improves both online and final holdout performance across four benchmarks under the shared-initialization continual protocol. The component and recurring-stream ablations further support the usefulness of expert management and knowledge consolidation.

Critical weaknesses
  1. The comparisons do not yet fully establish the extent of SEED’s advantage over existing adaptation methods. In Tables 1 and 2, several baselines substantially underperform Frozen, with some accuracy results approaching 50%. Although Appendix B.2 describes their architectural adaptations, the baseline hyperparameter-selection procedures and diagnostics for these large degradations remain insufficiently documented. Moreover, Table 3 retains method-specific initializations: several methods start from pretrained CLIP, whereas SEED uses a source-trained detector. Consequently, the large episodic performance gaps cannot be attributed solely to the adaptation mechanisms. Comparisons with closely related expert-based CTTA approaches, such as BECoTTA and Zhao et al. (2026), are also absent. These limitations reduce confidence in the breadth and attribution of the claimed comparative advantage.

  2. The stability of expert creation under dynamically changing class proportions remains insufficiently validated. DSEM detects changes in the marginal base-logit distribution, which can change when the real/generated proportion varies even if the class-conditional distributions remain unchanged. The decision-orthogonal projection is used for routing and does not directly remove this confounding factor from the expert-creation criterion. Table S2 evaluates different fixed proportions across separate runs, rather than changing proportions within a stream. Although the framework does not require experts to correspond one-to-one with generators, it remains unclear whether prior shifts alone cause redundant expert creation, repeated adaptation warm-up, or performance degradation. This uncertainty matters for the claimed suitability for continual deployment on non-stationary streams.

Questions for authors
  1. For Tables 1 and 2, could the authors clarify how baseline hyperparameters were selected, including the development data, learning rates, and update budgets? What implementation checks or diagnostics were performed to explain the substantial degradation of several methods relative to Frozen, such as predicted class proportions or adaptation trajectories?

  2. For Table 3, could the authors report each method’s performance before adaptation and its corresponding adaptation gain, given the different initializations described in Appendix B.2? Where compatible, a shared-initialization comparison would help isolate the contribution of adaptation. Please also explain whether closely related expert-based CTTA methods, such as BECoTTA and Zhao et al. (2026), can be evaluated under this protocol, or identify the specific obstacles to a meaningful comparison.

  3. When only the real/generated proportion changes within a stream, how does DSEM’s expert-creation mechanism behave? A controlled evaluation with a fixed generator and changing class proportions, reporting expert counts, creation events, routing behavior, and balanced accuracy, would directly address this concern. Do the existing results or additional analyses indicate whether such shifts lead to redundant experts or detection degradation?

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

SEED addresses continual test-time adaptation for AI-generated image detection without target-domain labels, generator identities, or explicit domain boundaries. It employs DSEM to manage a dynamic expert pool, AUE to generate pseudo-labels and estimate their reliability, and PKC to maintain feature statistics for balanced replay. Experiments on four public benchmarks evaluate online and final detection performance under diverse distribution shifts.

Critical strengths
  1. Existing AI-generated image detection methods mainly focus on generalization to unseen generators. The paper addresses the practical problem of continual test-time adaptation to evolving generators without labels or explicit domain boundaries.

  2. The overall design is well motivated. Freezing the base detector and maintaining lightweight residual experts allows context-specific adaptation while preserving previously acquired knowledge, balancing plasticity and stability.

  3. The ablation study supports the benefits of expert reuse. On the recurring stream SD1.5 → BigGAN → ADM → SD1.5, SEED achieves 82.44% final accuracy, compared with 75.60% without expert reuse, while reducing AUC forgetting from 4.33 pp to 0.09 pp.

Critical weaknesses
  1. Table 2 shows a substantial gap between the gains in Accuracy and AUC. Much of the Accuracy improvement may come from score recalibration rather than improved ranking ability. The authors should report accuracy at the optimal threshold for comparison.

  2. Although the Related Work section discusses closely related modular CTTA methods, such as BECoTTA[1] and Zhao et al.[2], these methods are not included in the experiments. The authors should compare SEED with strong baselines that incorporate dynamic expert mechanisms and are appropriately adapted to the task.

  3. The contribution of DSEM's orthogonal routing design is not independently evaluated through ablation studies.

  4. AUE uses a GMM to model base logits and assigns real/fake pseudo-labels based on component means. When the base detector struggles with a new generator, the two logit distributions may overlap substantially, making pseudo-labels unreliable even when their confidence weights are high.

[1] BeCoTTA: Input-dependent Online Blending of Experts for Continual Test-Time Adaptation. ICML 2024

[2] Shared & Domain Self-Adaptive Experts with Frequency-Aware Discrimination for Continual Test-Time Adaptation. AAAI 2026.

Questions for authors
  1. Although SEED does not store historical images, each new expert increases the number of residual MLP parameters and requires additional GMMs, routing prototypes, and feature statistics. The authors should report model size, online adaptation time, and inference latency as the expert pool grows to quantify the overhead of long-term deployment.

  2. Recent few-shot adaptation methods for AI-generated image detection include FiSeR[1], Fleet[2], and FSD[3]. Although these methods use a small number of labeled target-domain samples and are therefore not directly comparable to SEED's label-free setting, separate few-shot experiments would help clarify SEED's advantages without target labels and the practical value of its expert-management mechanism over simpler few-shot classifier adaptation.

  3. The main text contains numerous references to appendices, but the provided PDF does not include them.

[1] FiSeR: Fine-Grained Source Representations for Cross-Domain AI Image Detection. ICML 2026

[2] Fleet: Few Shots Lead Effective AI-generated Image Detection. ICML 2026.

[3] Few-Shot Learner Generalizes Across AI-Generated Image Detection. ICML 2025.

Reviewer_M9xvOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SEED for continuous test-time adaption in AI-generated image detection. The method freezes the base detector, dynamically creates residual experts based on changes in the base-score distribution, and employs decision-orthogonal features for expert routing. By combining pseudo-label estimation based on a Gaussian mixture model, reliability weighting, and statistical feature replay, it continuously updates the experts without requiring ground-truth labels, generator identities, or domain boundary information. The paper evaluates the detection performance on four public benchmarks and analyze the roles of individual components through ablation, repeated domain, and robustness experiments.

Critical strengths
  1. The method is designed with a clear motivation. Dynamic expert creation, parameter isolation, and reuse of historical states help balance adaptation to new distributions with knowledge retention. Distribution-driven pseudo-labels, reliability weighting, and statistical replay enable the utilization of unlabeled inputs and mitigate mismatches caused by fixed thresholds.

  2. The experiments cover four benchmarks and include module ablation, recurring-domain, and robustness analyses. Under the tested protocols, the method achieved better detection results compared to the selected baselines.

Critical weaknesses
  1. The value for real-world deployment has not yet been fully established. The main experiments were conducted with generators segmented sequentially, and scenarios involving the mixing of multiple generators within the same batch or variations in the true/false ratio were not sufficiently validated. Additionally, comparisons with strong static AIGI detectors under a unified protocol are lacking; therefore, the necessity and benefits of online adaptation under more practical input conditions remain undetermined.

  2. The reliability of the central mechanisms has not been sufficiently validated. The necessity and stability of orthogonal projection and mean-prototype routing have not been independently verified, and pseudo-supervision relies on the distinguishability of true and false scores. Analysis under conditions without projection, with random-direction projection, and with overlapping score distributions is needed to clarify the routing criteria, the boundaries of pseudo-label failure, and the impact of erroneous adaptation on historical detection performance.

Questions for authors
  1. The current experiments use consecutive generator-specific segments. Does SEED remain effective when real images and images from multiple generators arrive in a mixed stream?
  2. Under matched evaluation conditions, what additional benefits does SEED’s online adaptation provide over SOTA static AIGI detectors?
  3. Can targeted ablations of decision-orthogonal projection or prototype estimation further validate the reliability of expert routing?
  4. Can experiments using hard examples with strongly overlapping real–fake base-score distributions establish the failure boundary of pseudo-label-based adaptation?

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

Related papers

Discussion (0)

Sign in to comment.