Contradiction-Gated Revision for Deepfake Explanations

When a vision-language model explains why a face image is fake, it often makes up the evidence. It points to a "blending seam at the jaw" or "inconsistent lighting on the cheek" that is not there, and it does so most often when its real/fake verdict was a lucky guess. Asking the model to reread and fix its own explanation does not help, because models that check their own reasoning without outside evidence tend to invent new errors and turn right answers into wrong ones. We propose a different kind of revision. The explanation is broken into small claims, each naming a place and a type of artifact. Each claim is checked by forensic measurements that do not depend on the model: a separate segmentation or click-point detector, a frequency-noise score, and the consistency of reflections in the two eyes. The model is asked to revise only the claims these instruments contradict. Its verdict may change only when the evidence that survives checking no longer supports it.

The opportunity is practical. Newsrooms and courts need explanations whose individual statements can be audited, and the field already has benchmarks that score the truth of each claim, not only whether the verdict is right. The key design choice is to gate revision on outside contradiction instead of the model's own doubt. That choice addresses the failure that has made self-correction unreliable. On top of the gate, a calibration step picks acceptance thresholds so that the rate of unsupported claims in a final explanation has a stated upper bound.

Core Claim

A revision loop is safe and useful for forensic explanation under two conditions. First, it is triggered only by disagreement with independent measurements, never by the model's own reconsideration. Second, it acts on atomic claims that carry a location, not on free text. Under those conditions it should lower the rate of made-up evidence well below that of the strongest single-pass explainers. It should do this without the "correct-to-incorrect" verdict flips that ungated self-correction causes. It should also recover real artifacts that a verify-and-delete filter would simply lose.

That last clause is the falsifiable part. If deleting failed claims in one pass matches the loop on every metric, the loop is unnecessary. We treat that outcome as a live possibility and test it head-on.

Prior Work and Positioning

Three lines of work motivate the gap.

TriDF is a benchmark that scores deepfake explainers separately on perception, detection, and hallucination. It reports a correlation of about −0.60 between hallucination and detection accuracy, and it offers no way to reduce hallucination. Large Language Models Cannot Self-Correct Reasoning Yet shows that asking a model to check itself degrades accuracy on GSM8K: GPT-4 falls from 95.5% to 89.0% and Llama-2-70b from 62.0% to 36.5%. It concludes that correction needs an external signal. That result constrains our design: any trigger that comes from the model itself is ruled out. Do Language Models Know When They're Hallucinating References? shows that invented content becomes unstable when the model is questioned indirectly. We reuse that technique as one inexpensive checker.

Two works supply method pieces.

REFINER pairs a reasoning model with a separate critic that returns structured feedback of the form "error type, position, hint". It shows that a small critic can improve a frozen larger model. We adopt the structured-feedback format, but our critic is a set of measuring instruments instead of a trained language model, and it works on images instead of math. Mitigating LVLM Hallucination at Post-hoc via Multi-agent System routes image descriptions through a supervisor, an object detector, and two debating agents. It checks only whether everyday objects exist. Forensic claims are harder because they concern the absence of physical consistency, not the presence of an object.

The closest forensic systems all work in a single pass.

REVEAL combines eight expert signals into one chain of evidence and reports 95.00% mean accuracy. Multi-Agent Forensic Reasoning has four agents report to a judge in one round, reaching 53.28% out-of-distribution F1 against 47.45% for the best single model. Fake-in-Facext interleaves text with mask tokens to reach 30.4 mIoU on artifact segmentation, but it never checks whether a masked region is actually anomalous. PRPO trains paragraph-level rewards for visual and prediction consistency. It reaches 89.91% average F1 on DF-R5, and its test-time adaptation updates weights through reinforcement learning.

Our proposal differs from all of these in two ways. It revises specific claims at inference time in response to named contradictions, and it gives a calibrated bound on the claim error rate. Distribution-free bounds of this kind have been used for question answering in TRAC, but not for forensic explanations.

Research Program

The generator is an open vision-language model that can point to regions. We use FiFa-MLLM, which emits masks, and Qwen2.5-VL-7B, which we prompt to emit boxes. Its first-pass explanation is parsed into claims. Each claim is a triple of region, artifact concept, and polarity, with the concept drawn from Fake-in-Facext's facial concept tree.

Each claim is sent to a checker matched to its concept:

  • Texture and blending claims are scored by a separately trained region anomaly detector and a high-pass noise statistic.
  • Eye-lighting claims are scored by the corneal-highlight IoU test from Exposing GAN-generated Faces Using Inconsistent Corneal Specular Highlights. On frontal portraits this test separates real faces (IoU typically above 0.5) from StyleGAN2 faces (often below 0.4).
  • Video location claims are compared with the ExDDV click predictor.
  • Every claim also receives an indirect-query consistency score. We sample several answers to related questions about the region, such as the light direction or which neighboring structures are present, and measure how much they disagree.

A checker returns "supported", "contradicted", or "out of scope". For example, the corneal test abstains on non-frontal faces. Only a "contradicted" result produces feedback, formatted REFINER-style with the claim identity, the failing measurement, and the region. The generator then gets one targeted turn per contradicted claim. In that turn it may retract the claim, relocate it, or name a different artifact at the same region, and any new claim must pass the checkers in turn.

Two further rules govern the loop:

  • Verdict rule. A "fake" verdict must rest on at least one surviving supported claim. Otherwise the system re-inspects regions where the checkers fired but the model said nothing, or it abstains.
  • Budget. The loop stops after at most three rounds.

Calibration uses held-out images with ground-truth artifact masks from FiFa-Bench and the TriDF perception annotations. On this set we choose acceptance thresholds with conformal risk control. The goal is that the expected fraction of accepted claims with no matching ground-truth artifact stays below a target $\alpha$. This bound holds only under the assumption that calibration and test data come from the same distribution. We therefore report separately how far the bound is violated on unseen generators, and we do not claim it holds there.

The analysis also measures how independent the checkers really are. If a checker was trained on the same data as the generator, its errors may be correlated with the generator's. We estimate how often a checker "supports" a claim that the ground truth rejects, and we report results with each checker family removed in turn.

Baselines

Do-nothing baseline. The generator's first-pass explanation and verdict, with no checking, from the same model under the same prompt. This is the reference for every change we report.

Strongest single-pass explainers.

  • PRPO on DF-R5 is the strongest competitor because it already rewards visual grounding of each paragraph. It reports 89.91% average F1 and 4.55/5.0 explanation quality (PRPO). We run the released model on the same DF-R5 test split. Our loop wraps a generator of comparable size, and we also wrap PRPO itself as the generator. That way the comparison asks whether gated revision adds anything on top of the best trained explainer.
  • Fake-in-Facext for localization. The reported number is 30.4 artifact-segmentation mIoU.
  • TriDF hallucination scores. No reported result was found on Lacuna for per-model TriDF hallucination scores, so we compute them ourselves for all systems with the benchmark's protocol.
  • Video. The comparison point is the multi-agent system's 53.28% out-of-distribution F1 on FaceVid-Forensics-100K.

Beating the baselines. The loop beats them if it lowers the TriDF hallucination score at an equal or better detection F1. It must stay within 1 point of PRPO's DF-R5 F1, and it must not reduce Fake-in-Facext's mIoU.

Compute-matched baselines and ablations.

  • Self-consistency. The same number of model calls spent on sampling explanations and keeping only claims that recur in a majority of samples. This matches the compute budget of the loop.
  • Verify-and-delete. The same checkers applied once, with failed claims deleted and no revision or re-inspection. This isolates the value of iteration.
  • Ungated self-refinement. The model critiques and revises its own explanation with the same number of rounds but no checker input. This reproduces the Huang et al. setting.
  • Random gating. Revision is triggered for a random subset of claims of the same size as the contradicted set. This shows whether the checker signal, and not just the extra turns, drives the gains.

All systems use the same generator weights, the same prompt-tuning budget on a shared development split, and the same calibration data.

Why not just…?

Why not just use self-consistency voting over sampled explanations? On GSM8K, majority voting over nine samples scored 88.2% against 83.0% for debate at equal calls (Huang et al.), so this is a serious competitor for verdict accuracy. For explanations, though, it fails in a specific way. TriDF finds that invented evidence tracks lucky guesses, and a lucky guess can recur systematically across samples, so frequent claims are not necessarily true. Our compute-matched self-consistency baseline tests this directly. If it matches our hallucination reduction, the checkers are adding nothing and we will report that.

Why not just check claims once and delete the failures? This is the most serious threat, and it may capture most of the reduction in hallucination. What it cannot do is replace a deleted claim with the real artifact nearby, or notice regions that the checkers flag but the model ignored. As a result it should lose artifact coverage, and on hard fakes it should push verdicts toward "real", which is PRPO's reported failure on flawless forgeries. The verify-and-delete ablation measures exactly the coverage and recall-of-fakes difference. If that difference is within noise, we will narrow the contribution to the calibrated deletion filter and drop the loop.

Why not just train a better single-pass explainer, as in REVEAL or PRPO? Training reduces hallucination on average but cannot correct a specific wrong claim at inference time. It also provides no bound on the claim error rate that holds for each deployment. Wrapping PRPO itself as the generator answers the question empirically: gains on top of PRPO show that the two approaches complement each other, and no gain shows that training already captures the benefit.

Decisive Checks

The first check is the flip audit. On DF-R5 and the TriDF image split, we count verdicts that go from correct to incorrect under each revision method. Ungated self-refinement should reproduce the Huang et al. pattern of many such flips. The gated loop should keep them near zero while fixing more incorrect-to-correct cases than it breaks. If the gated loop flips correct verdicts at a rate comparable to ungated refinement, the central claim fails.

The second check compares the calibrated claim error rate at $\alpha = 0.1$ on in-distribution generators against the rate on held-out generators: SiT from DF-R5 and the 20 unseen FaceVid generators. This shows how much of the guarantee survives a change of generator.

The third check reports artifact coverage and fake recall against verify-and-delete, separately for image and video. The video split is included because TriDF documents a roughly 50% drop in artifact coverage for current models when moving from images to video.

Finally, we report tokens and wall-clock cost per image for every system, so any accuracy gain can be weighed against the three- to ten-fold inference overhead typical of multi-turn refinement.