Teaching Deepfake Reasoners to See Before They Think

Detectors for AI-generated images increasingly work like a detective writing a report. A multimodal language model, meaning a language model that also reads images, looks at a picture, writes out step-by-step reasoning about lighting, geometry and texture, and then gives a verdict of real or fake. Writing reasoning before answering is called chain-of-thought. These systems now beat both classic classifiers and frontier chatbots on hard benchmarks. Every one of them, however, is built on a general-purpose vision encoder, the part of the model that turns pixels into the tokens the language model reads, and that encoder was pre-trained to describe everyday scenes. We suspect that when a new generator appears, the reasoning chain can only talk about evidence the encoder already encodes. Once the encoder stops registering a generator's traces, more deliberation mostly produces fluent justification of a guess.

We propose to test this directly and to fix it cheaply. Before any reasoning training, we continue pre-training only the vision encoder of a compact 7B multimodal model on a forensic self-supervised objective. The objective uses large unlabeled pools of real and generated images, and it scrambles the image's meaningful content so the encoder has to attend to generation traces instead of object semantics. We then apply the standard chain-of-thought and reinforcement-learning recipe on top. The experiments use a chronological split: train on older generators, test on newer ones. They are designed to show whether the benefit of chain-of-thought depends on what the encoder perceives, and whether forensic perception combined with deliberation beats either one alone.

Core Claim

Our claim is that in reasoning-based forgery detectors, generalization to newer generators is limited by perception, not deliberation. If so, the gain from chain-of-thought should interact with encoder quality. It should be small, or even harmful because it adds rationalization, on a generic encoder facing unseen generators. It should be larger on a forensically pre-trained encoder. This also resolves a live contradiction in the literature. Beyond Semantic Features: Pixel-level Mapping shows that scrambling adjacent-pixel statistics, which destroys semantic content, forces a plain ResNet-50 onto generation traces and reaches 98.4% on GenImage. In contrast, Veritas and C2P-CLIP argue that low-level traces are temporary and that robust detection needs semantic, physical reasoning. Our position is that the two belong at different layers. Semantics should be removed from the encoder's pre-training signal, and semantic reasoning should be kept in the language model's deliberation.

Relation to Prior Work

The closest work consists of reasoning detectors that specialize a general multimodal model with supervised traces and then reinforcement learning. Veritas teaches a five-stage forensic reasoning pattern and penalizes "overthinking." It reaches 90.7% average accuracy on HydraFake. VidGuard-R1 does the same for video, with rewards for detecting temporal glitches and for estimating how far a video has progressed through denoising. AIGI-Holmes comes nearest to our idea: it adds a pre-trained low-level visual expert beside LLaVA's encoder and fuses the expert's prediction at decoding time.

Our proposal differs from all three in two ways. First, we change the base encoder itself through continued pre-training on unlabeled forensic data, rather than attaching an expert or relying on fine-tuning alone. Second, we measure the interaction between perception and deliberation instead of reporting a single end-to-end system.

A second group of detectors wraps external tools around the language model. REVEAL feeds the outputs of eight frozen forensic expert models into the language model. EDVD-LLaMA relies on facial-landmark tools. The multi-agent video system runs four 7B agents in parallel. These works motivate our question, because their language model never perceives forensic signal directly. They also supply a strong baseline, since "expert as a text tool" is the cheap alternative to pre-training.

Towards Universal Fake Image Detectors shows that frozen CLIP features generalize well. It also identifies the "sink class" failure, in which "real" simply means "no artifacts I have seen before." This constrains our objective design: pre-training must model real images positively, not only learn to spot fakes. Specializing Smaller Language Models towards Multi-Step Reasoning finds that specializing a compact model destroys its general reasoning. This constrains us further, and it is why we pre-train only the vision tower and measure retention. Finally, domain pre-training results in other fields, such as MonoCoder and the PEFT code study, found that the base representation matters more than the fine-tuning method. They motivate the hypothesis but do not establish it for vision forensics.

Research Program

We start from Qwen2.5-VL-7B, the backbone used by VidGuard-R1 and the multi-agent system. Only its vision encoder receives continued pre-training. The corpus is unlabeled or weakly labeled: real photographs paired with images from generators released before a fixed cutoff date, drawn from the generator families in UniversalFakeDetect and GenImage, including GANs and SD v1.4/1.5.

The pre-training objective has three terms. The first is contrastive agreement between two pixel-mapped views of the same image, following the mapping of the pixel-level mapping paper. The mapping removes semantic content while keeping generation traces, so agreement must rest on those traces. The second is prediction of generator family, together with prediction of the number of denoising steps for diffusion samples. This reuses VidGuard-R1's generation-progress signal as a pre-training target instead of a reward. The third is a real-image reconstruction term that gives "real" a positive model, which addresses the sink-class problem. To protect the encoder's general scene understanding, a distillation term keeps its outputs close to the original encoder on unmodified real images.

On top of each encoder, generic or forensic, we train two heads under identical data and compute. One is a direct-verdict head. The other is a Veritas-style chain-of-thought model, trained with supervised reasoning traces followed by group relative policy optimization (GRPO), a reinforcement-learning method that scores each sampled answer against the other answers sampled for the same input. This produces a two-by-two design. Its interaction term is the scientific result.

We also run a faithfulness probe. At test time we replace or truncate the reasoning chain and measure how often the verdict changes. We also ablate the visual tokens that the chain cites and check whether the verdict follows. If reasoning is post-hoc rationalization, a problem REVEAL raises, verdicts should be insensitive to these interventions.

Baselines

The do-nothing reference is the untuned Qwen2.5-VL-7B prompted with a forensic instruction. For context we also report the frontier-model gaps given in Veritas: it exceeds Gemini-2.5-Pro by 11.8% and GPT-4o by over 30% on HydraFake.

The strongest existing method on our question is the Veritas recipe on a generic encoder. It reports 90.7% average accuracy, 90.3% cross-forgery and 82.2% cross-domain on HydraFake. We reproduce the recipe on our backbone. To make the comparison fair, the generic-encoder arm receives the GPU-hours we spend on encoder pre-training as additional supervised and reinforcement-learning steps on the same pre-cutoff generators. Beating it means higher accuracy on the post-cutoff and cross-domain splits at matched compute. We set the threshold at the reproduced Veritas cross-domain figure plus the seed-to-seed spread across three runs.

The strongest non-language competitor is the fixed pixel-mapping ResNet-50. It reports 98.4% on GenImage, 9.6% above UniFD and 2.6% above C2P-CLIP. We retrain it on our pre-cutoff pool. For CLIP-family detectors we include C2P-CLIP, which reports 93.8% mean accuracy on UniversalFakeDetect, and a linear probe on frozen CLIP. Both are trained on the same pre-cutoff data.

The tool-augmentation baseline gives the generic-encoder chain-of-thought model the pixel-mapping classifier's score and residual map as text input, in the style of REVEAL. It is the cheapest way to hand forensic perception to the reasoner. The key ablation isolates semantic scrambling: the same encoder pre-training without pixel-mapped views, using ordinary augmentations only.

Why not just…?

Why not just use the pixel-mapping classifier? This is the most serious threat. A ResNet-50 that reaches 98.4% on GenImage might match any 7B reasoner on accuracy. The file reports no result for it on HydraFake, on DFBench's newer generators, on partial edits or on RAID. Our chronological split settles this. If the classifier equals our model on post-cutoff generators and on real images with natural distortions, then the accuracy claim fails, and our contribution shrinks to faithful explanations. We would report that honestly. We expect it to lose where evidence is semantic or partial. DFBench reports that full generations are detected well but localized AI edits are much harder. DFBench also documents a "distortion trap," in which low-level detectors misread natural blur and compression as signs of forgery.

Why not just prompt or fine-tune harder on the generic model? The Veritas recipe with extra compute is exactly our matched-compute baseline. If extra reinforcement learning closes the gap, perception was not the bottleneck and our claim is wrong. The two-by-two interaction answers this directly.

Why not just feed forensic expert outputs to the language model as text? This is REVEAL's approach, and it costs nothing in pre-training. Its weakness is that when the frozen expert fails on a new generator, the reasoning built on it fails too. We compare it against our model on the post-cutoff split. If it matches, tool use is the better engineering choice and we will say so.

Why not just use frozen CLIP features? RAID shows that CLIP-based and commercial detectors collapse under small adversarial perturbations: Corvi et al.'s black-box AUROC falls from 0.99 to 0.67, and Sightengine's accuracy drops from 0.66 to 0.0. A frozen CLIP detector also provides no rationale. We nevertheless include it at full strength.

Proposed Experiments

The primary test is the chronological split. We train on pre-cutoff generators and evaluate on post-cutoff generators from DFBench, including SD3.5-Large, FLUX, PixArt-sigma, Janus and Infinity. DFBench names SD3.5-Large the hardest generator to detect but reports no number for it. We also evaluate on HydraFake's four out-of-distribution levels and on GenImage and UniversalFakeDetect for comparability. Metrics are accuracy and AUROC, with real-image accuracy on DFBench's distorted real subset reported separately; MoA-DF reports 98.41% there. Partial-edit accuracy is reported on DFBench's AI-edited subset. Robustness is measured by AUROC on RAID at all three perturbation budgets. The file reports no result on RAID for any chain-of-thought detector, so this measurement is itself new.

The decisive check is the interaction term: the chain-of-thought gain on a forensic encoder minus the chain-of-thought gain on a generic encoder, measured on post-cutoff generators with seed confidence intervals. A positive interaction supports the claim. A null interaction with a strong forensic-encoder main effect means perception matters but reasoning adds nothing to accuracy. A null result on both refutes the proposal.

The faithfulness probe must show that verdicts move under chain intervention more often with the forensic encoder. If they do not, better accuracy cannot be credited to reasoning.

Two further checks guard against hidden costs. Retention of general multimodal ability is measured on VeriSciQA, because the multi-step specialization paper found large general-capability losses after narrow specialization. The semantic-scrambling ablation must lose post-cutoff accuracy. Otherwise the reconciliation of the semantics-disruption and semantic-reasoning positions described above is unsupported.