acceptodds
Withdrawn from ICLR 2027

CLARE: Continuous Latent Reasoning for AIGC Image Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

As AI-Generated Content (AIGC) advances rapidly, the growing risk of its malicious exploitation creates an urgent need for reliable AIGC image detection. While recent approaches based on Multimodal Large Language Models (MLLMs) have achieved promising detection performance, their reliance on explicit textual chain-of-thought (CoT) introduces two fundamental limitations. First, autoregressively decoding continuous representations into discrete text tokens creates an information bottleneck, compromising rich visual cues. Second, as generated images become increasingly realistic, constructing high-quality and hallucination-free textual reasoning chains becomes substantially more difficult. To overcome these limitations, we propose CLARE, a Continuous Latent Reasoning framework for AIGC image detection. CLARE conducts forensic reasoning entirely within the continuous latent space of MLLMs by directly feeding the last hidden state of each step as the input representation for the next step, thereby bypassing discrete text token generation. Furthermore, we pretrain a task-specific vision encoder on AIGC detection and inject its representations into the latent reasoning process, effectively equipping the general multimodal backbone with forensic sensitivity. Extensive experiments across multiple challenging benchmarks demonstrate that CLARE achieves superior detection performance and exhibits strong cross-generator generalization.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AIGC image detection

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

The paper proposes CLARE for AI-generated image detection. It combines continuous latent computation, regional visual alignment, and features from a separately trained forensic encoder, removing the need for textual chain-of-thought annotations. Evaluation covers five benchmarks and four Qwen backbones. The reported results reach 99.44% accuracy on Holmes-Set and 99.50% on AIGI-Now, with additional experiments on recent generators, component ablations, and image post-processing.

Critical strengths
  1. Strong results across several benchmarks. CLARE performs consistently well across diverse generators and forgery types. Per-generator results and post-processing experiments provide useful evidence beyond aggregate accuracy.

  2. A practical framework with useful ablations. The method uses authenticity labels without requiring textual reasoning annotations. The component and backbone experiments support the usefulness of combining regional alignment with forensic features.

Critical weaknesses
  1. The reasoning comparison is confounded by additional components. Table 5 compares full CLARE with textual CoT, while CLARE also receives regional alignment supervision and a separately trained forensic encoder. Table 7 reports 93.29% average accuracy for latent-only reasoning, compared with 93.05% for textual CoT; alignment and forensic injection raise this to 97.67%. Matched baselines receiving the same forensic features are needed to establish the benefit attributable to latent reasoning.

  2. Cross-generator detection is only partly validated. T2I-CoReBench and MNW contain only generated images, so their reported accuracy measures fake recall. False-positive behavior on corresponding real-image distributions remains unclear. The exact training splits and generator overlap across both training stages and evaluation benchmarks also need specification to substantiate the unseen-generator claim.

Questions for authors
  1. Do the answer-only and textual CoT baselines receive the same forensic features as CLARE? Can you provide a comparison that holds this information constant?

  2. What are the false-positive rates on real images under the cross-generator evaluation, using a threshold fixed before testing?

  3. Please specify the training splits and generators used in both stages, and their overlap with each evaluation benchmark.

Reviewer_uT4XOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes CLARE, a continuous latent reasoning framework for AIGC image detection. The paper argues that conventional textual CoT repeatedly decodes high-dimensional continuous hidden states into discrete tokens and then feeds the corresponding token embeddings into the next step. This decode-and-embed process may discard fine-grained forensic evidence that is difficult to verbalize. CLARE therefore directly uses the hidden state from each step as the input to the next over a fixed sequence of K reasoning steps, producing the final real/fake prediction only after latent reasoning is completed.

Building on the basic Latent Only variant, the authors introduce two additional visual mechanisms. Visual alignment supervises the latent hidden states with regional representations from the original frozen vision encoder, associating different reasoning steps with different image regions. Forensic injection first trains a task-specific forensic encoder using real/fake classification and then injects its regional forensic representations into the corresponding latent reasoning steps. The full model thus combines continuous hidden-state propagation, regional visual alignment, and task-specific forensic feature injection.

Critical strengths

The paper is clearly written overall. The research motivation, two-stage training pipeline, and organization of the main modules are relatively complete, and the figures intuitively illustrate the distinction between Textual CoT and Continuous Latent Reasoning.

The experiments cover multiple datasets, generator types, and forgery granularities, and provide analyses from several perspectives, including backbone architectures, numbers of reasoning steps, model components, and robustness to image post-processing.

Keeping intermediate reasoning in a continuous latent space and avoiding the forced verbalization of fine-grained forensic information is an interesting idea worthy of further investigation. However, the independent contribution of the continuous latent-space mechanism still needs to be demonstrated through more rigorous and properly matched controlled experiments.

Critical weaknesses

This paper positions CLARE, a Continuous Latent Reasoning framework, as its core innovation and attributes the performance improvements to continuous latent reasoning. However, in my view, the current experiments are insufficient to support this central conclusion.

First, the paper claims that autoregressive text generation maps high-dimensional continuous hidden states to discrete vocabulary tokens, and that this decode-and-embed process causes information loss and limits model performance. This claim serves as the primary motivation for introducing continuous latent reasoning. If so, the most direct and critical experiment should compare continuous and discrete intermediate-state propagation under fully matched conditions.

Specifically, the Latent Only setting should be compared with a variant that uses exactly discrete intermediate tokens, i.e., standard textual decoding. This discrete variant need not apply natural-language CoT supervision to the intermediate tokens, nor must these tokens have human-readable semantics. It should simply project the hidden state at each step onto the discrete vocabulary and feed the corresponding token embedding into the next step. This would ensure that: (1) the two methods have the same number of intermediate reasoning tokens and approximately matched computational costs; and (2) neither method relies on human-annotated textual reasoning to supervise the intermediate reasoning process.

If this length-matched discrete textual CoT, trained with the same supervision scheme, achieves performance comparable to Latent Only, then the improvement of Latent Only over Answer-only would more likely arise primarily from the additional computation rather than from the claimed advantage of the “continuous latent space.”

The Textual CoT baseline in the paper cannot substitute for the experiment described above. This baseline explicitly supervises the intermediate reasoning text, and its performance may therefore depend on the quality of the Holmes-Set reasoning annotations, the redundancy of the explanations, and potential annotation errors. Moreover, Textual CoT generates 590 tokens on average, whereas CLARE contains only 36 latent reasoning steps. The two are therefore clearly mismatched in both reasoning length and computational cost.

In fact, even under the paper’s own Textual CoT comparison protocol, cross-referencing Tables 5 and 7 shows that Latent Only has no substantial advantage over Textual CoT. Their average accuracies are 93.29% and 93.05%, respectively, a difference of only 0.24 percentage points. Furthermore, Latent Only performs worse on three of the five datasets.

Second, according to the existing ablation results, the main gains of the full CLARE model over Latent Only actually come from Visual alignment and Forensic injection. Visual alignment uses regional representations from the original frozen vision encoder to supervise the latent hidden states. Forensic injection introduces a task-specific forensic encoder pretrained with real/fake classification supervision and directly injects its regional representations into every latent reasoning step.

However, the paper does not directly demonstrate that the full CLARE model outperforms the following two critical baselines:

  1. The forensic vision encoder trained in Stage 1, directly connected to its binary classification head and used as a standalone forgery detector.
  2. An MLLM trained by directly concatenating or simply fusing the representations from the original vision encoder and the task-specific forensic encoder, without multi-step continuous latent reasoning.

These two baselines are essential for determining whether continuous latent reasoning plays an irreplaceable role, yet neither is provided. What the current experiments actually demonstrate is that, when the original vision encoder is not adapted to the forgery-detection task, adding a dedicated forensic encoder trained with real/fake supervision and using its representations through regional alignment and step-wise injection improves detection performance. This result is entirely expected, but it is helpless for demonstrating the effectiveness of the continuous latent reasoning mechanism.

In other words, without the above baselines, CLARE can be viewed as a relatively complex and elaborate visual feature alignment and injection method combined with the “latent reasoning,” rather than as a sufficiently validated reasoning mechanism. The paper may certainly claim that its regional alignment and forensic feature injection outperform more conventional and direct feature-fusion approaches. However, to support such a claim, it should at least include the two baselines above and conduct comparisons under matched supervision, task-specific visual information, parameter counts, and computational costs.

Questions for authors

It is necessary to address the key issues identified under "Critical Weaknesses."

Reviewer_TC8tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a continuous latent reasoning framework (i.e., CLARE) for AIGC image detection. It detects AI-generated images through continuous latent reasoning, combining step-wise visual alignment with region-level forensic features. It achieves strong performance across benchmarks, generators, and MLLM backbones while using substantially fewer tokens than textual CoT.

Critical strengths
  1. Continuous latent reasoning may preserve fine-grained forensic cues lost in textual CoT.
  2. Extensive evaluations across several datasets and baselines demonstrate competitive detection performance. Compared with textual CoT, CLARE achieves higher accuracy while generating fewer tokens, offering a favorable balance between detection effectiveness and output length.
Critical weaknesses
  1. The claimed discretization-induced information bottleneck is not directly validated. Continuous latent reasoning alone achieves only a marginal improvement over textual CoT (93.29% vs. 93.05%), while larger gains arise from visual alignment and forensic feature injection. Thus, the extent to which avoiding textual discretization contributes to the final performance remains unclear.

  2. The methodological distinction from Deep-VRM requires stronger empirical justification. Both methods inject artifact-sensitive forensic features into MLLM representations, whereas CLARE introduces region-specific injection at each latent step and step-wise visual alignment. Controlled comparisons with a Deep-VRM-style single-boundary injection variant would help disentangle these contributions and establish the advantages of CLARE’s design.

  3. Baseline configurations are insufficiently documented, limiting the comparability of the reported results. In particular, Table 1 directly adopts Deep-VRM’s original results with Qwen2.5-VL-7B, whereas CLARE uses Qwen3-8B-VL by default. Comparisons using a shared backbone and evaluation pipeline would strengthen the empirical conclusions.

  4. The necessity of the dual-encoder design is insufficiently established. Although the forensic encoder supplies artifact-sensitive features, latent states are aligned with representations from the native encoder. Comparisons among forensic-only, native-only, and dual-encoder variants would clarify the benefits of this choice and its computational and parameter overhead.

  5. The fixed correspondence between latent reasoning steps and image regions lacks empirical justification. The paper adopts a left-to-right, top-to-bottom scanning order without examining alternative orderings. Evaluating different region sequences and alignment strategies would clarify the method’s sensitivity to these design choices.

Questions for authors

Please see the weakness section.

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

Related papers

Discussion (0)

Sign in to comment.