Catching Style-Camouflaged Entity Swaps by Reading Multilingual Coverage Through a Delta-Rule Memory

Fake-news detectors learned to spot sensational, sloppy writing, and attackers stopped writing that way. An LLM can take a false story and rewrite it in calm wire-service prose, which cuts automated detectors' accuracy by more than 20% Can LLM-Generated Misinformation Be Detected?. The hardest version keeps a real article intact and swaps only one name, so every stylistic cue is genuine. Here the best knowledge-graph detector reaches only about 67–68% accuracy Automatic Detection of Entity-Manipulated Text using Factual Knowledge. Any signal that lives in the prose can be removed by a better rewrite. A signal that lives in the facts is harder to remove. When a real event happens, dozens of outlets in several languages cover it, and a swapped name in one article disagrees with that coverage however the article is phrased.

We propose a detector that reads the multilingual coverage of an event as one long stream. It stores who-did-what bindings in a fixed-size memory, then measures how badly each entity in a target article violates that memory. The memory is a delta-rule linear recurrence, the DeltaNet layer. It is a recurrent model whose state is a matrix updated by subtracting its own prediction error, and it can be trained in parallel over sequence length Parallelizing Linear Transformers with the Delta Rule over Sequence Length. The delta rule already computes a per-token "surprise" (the prediction error), and we use that surprise directly as the detection signal. The recurrent form matters for deployment as well as speed. A newsroom can keep one small state per developing story and check each new article against it in time proportional to that article's length, without re-reading the whole bundle.

Core Claim

Our claim is that manipulation evidence defined as a binding conflict between a target article and cross-lingual coverage is invariant to register (the tone and style of the writing) in a way that stylometric evidence is not. We further claim that a delta-rule memory is the right mechanism for extracting that evidence at the scale of a real event bundle. The claim is falsifiable in three ways. Accuracy of our detector should stay nearly flat when the attacker adds LLM restyling and cross-lingual rewriting on top of the swap, while style-sensitive detectors drop. A matched recurrent memory without error-driven overwriting (Mamba or Gated Linear Attention) should do measurably worse on single-entity swaps. At equal accuracy, cost per new article should stay constant as the bundle grows, whereas an attention model's cost should keep rising.

Each input direction plays a different role. Adversarial stylistic mimicry defines the threat. Cross-lingual style transfer contributes two things. It supplies the attack, because the swapped article is restyled and re-voiced into another language's journalistic register, as BLUFF's generation pipeline does at scale BLUFF. It also supplies a training method: the dual contrastive heads that handwriting models use to make a representation invariant to writer style but sensitive to character identity Disentangling Writer and Character Styles for Handwriting Generation. The sequential processing bottleneck, meaning the problem that recurrent models must process a sequence one step at a time, is both the constraint and the reason to use recurrence at all.

Research Program

The camouflaged-swap benchmark. We first build the benchmark, because none exists. We start from same-event article clusters. CrossNews-UA, a cross-lingual news similarity benchmark across Ukrainian, Polish, Russian and English, supplies aligned multilingual coverage. The human fact-checked portion of BLUFF supplies breadth across languages. RealNews supplies English clusters, which we form by date plus entity-overlap retrieval so we can compare against Jawahar et al.

For each cluster we hold out one target article and apply four attack levels. The first is a one-, two- or three-entity swap following Jawahar et al.'s protocol. The second adds a swap followed by an LLM rewrite into neutral authoritative register, the controllable-misinformation setting of Chen et al. The third adds a swap followed by translation and restyling into the register of an outlet in another language. The fourth is a "two-hop" swap, in which the swapped entity is used consistently throughout the target article, so the conflict with the bundle appears only through a chain of two relations. Clean restyled articles are included as negatives, so that any detector that flags restyling itself is penalized.

Model. A multilingual token embedding feeds a stack of DeltaNet layers trained with the chunkwise algorithm, which processes blocks of tokens in parallel and passes only the memory state between blocks DeltaNet. The bundle is read first, and the target article is read last. In each layer the memory $S_t$ (a matrix) is updated from a key $k_t$, a value $v_t$ and a write strength $\beta_t \in (0,1)$:

$$S_t = S_{t-1} + \beta_t\,(v_t - S_{t-1}k_t)\,k_t^\top, \qquad e_t = \lVert v_t - S_{t-1}k_t \rVert .$$

The quantity $e_t$ is the memory's prediction error at token $t$. We pool $e_t$ over entity-mention tokens of the target article to get a document score, and take the argmax to localize the swap. A small classification head reading the final state is trained jointly. The ablations compare it against the raw error score.

Register invariance. We borrow the handwriting models' two-head contrastive recipe. One head is trained to pull an article together with its own restyled and translated versions and push away its swapped versions, which makes keys and values carry content rather than register. A second head does the opposite and is discarded at test time. This mirrors how writer and character styles are separated in handwriting generation, and it is the one ingredient we expect to matter most under the heaviest restyling attacks.

Streaming. At inference, each event keeps its state $S$. A new article is scored against the frozen state and then written into it. Per-article cost is therefore independent of how many articles came before.

Baselines

The naive reference point is a fine-tuned mDeBERTa classifier on the target article alone, with no evidence. BLUFF reports that mDeBERTa reaches the highest overall F1 on its external datasets at 67.3%, with a 9.9-point macro-F1 gap between high- and low-resource languages under cross-lingual transfer BLUFF. We expect this model to do well on unrestyled swaps only where style leaks and to degrade under restyling. Beating it means a smaller drop between attack levels one and three, not just a higher score at level one.

The strongest existing method for this exact threat is the RoBERTa-plus-graph-network detector grounded in the YAGO 4 knowledge base. It reports roughly 67–68% accuracy on single-entity swaps, roughly 84% on three-entity swaps, and 70–80% localization precision on RealNews Automatic Detection of Entity-Manipulated Text using Factual Knowledge. We rerun it on our RealNews-derived split using its original knowledge base, and we additionally give it access to our bundles by linking bundle entities into its graph, so that it is not disadvantaged by lacking evidence. Beating it means exceeding 68% on single swaps on the same split, and holding that margin on the restyled and cross-lingual attack levels. On those levels we expect its text branch to be unaffected but its coverage of low-resource entities to fall.

The strongest style-robustness baseline is AdStyle, an iterative LLM loop that generates adversarial restyling prompts and trains the detector on the outputs. It reports AUC 0.9174 on PolitiFact under attack, against 0.8127 for a vanilla detector Adversarial Style Augmentation via Large Language Model for Robust Fake News Detection. We apply AdStyle augmentation to the mDeBERTa baseline with the same LLM budget we spend building our own restyled training data.

The strongest evidence-using baseline is a pairwise cross-encoder: mDeBERTa scores the target against each bundle article, and the scores are then aggregated. This is the neural analogue of the cross-lingual evidence feature of Cross-lingual Evidence Improves Monolingual Fake News Detection, for which no reported result found on Lacuna for its combined CE-plus-BERT accuracy. We also run a long-context Transformer that reads the whole bundle with full attention.

All models get the same bundles, the same training attacks, parameter counts within 10% of each other, and the same hyperparameter search budget.

Two ablations isolate the key ingredient. The first swaps DeltaNet for Mamba and for Gated Linear Attention at matched state size and depth, keeping everything else fixed. This tests whether error-driven overwriting is what catches single swaps, as the associative recall results in the DeltaNet paper would predict. The second removes the register-contrastive heads. Two controls check that the model actually uses the evidence: shuffling the bundle to an unrelated event, and deleting it.

Why not just…?

Why not just use a long-context Transformer or a pairwise cross-encoder over the same evidence? This is the real threat. Nothing in the file shows that attention fails at within-bundle binding checks, and full attention is at least as expressive as DeltaNet for this task. We expect it to match us on accuracy at small bundle sizes. The proposal stands only if one of two things holds. Either the delta-rule error is a better localized signal at equal accuracy, or the Transformer's cost per new article grows with bundle size while ours stays flat. The streaming experiment below settles the second point directly. If attention wins on accuracy by a clear margin at every bundle size we test, the contribution shrinks to an efficiency trade-off, and we will report it that way.

Why not just harden a single-article detector with adversarial style augmentation? AdStyle reaches strong AUC on English, but only on English short claims, and the augmented detector still reads nothing except the target article AdStyle. An entity swap on authentic prose leaves no style trace to become robust to. That is why stylometry fails on this attack A Stylometric Inquiry into Hyperpartisan and Fake News; Automatic Detection of Entity-Manipulated Text using Factual Knowledge. Our level-one results test this prediction directly. AdStyle should be near chance on unrestyled single swaps, and if it is not, our premise is wrong.

Why not just verify entities against a knowledge graph? A knowledge graph checks against slow-moving world knowledge, while bundle coverage checks against what was reported about this specific event. Jawahar et al. acknowledge degradation for novel and obscure entities, which describes most low-resource-language news. We report results split by whether the swapped entity exists in YAGO 4. We expect the two sources of evidence to be complementary, and we will test a combined model rather than claim replacement.

Why not just translate everything to English and run an English detector? The cross-lingual evidence study found that monolingual English evidence underperforms native cross-lingual evidence Cross-lingual Evidence Improves Monolingual Fake News Detection, and translation is itself one of the attack channels in our benchmark. We include a translate-then-detect pipeline as a baseline on attack level three.

Proposed Experiments and Decisive Checks

The primary result is detection accuracy and localization precision across the four attack levels and one-, two- and three-entity swaps, broken down by language resource tier. The decisive signature for the core claim is a flat curve for our model from level one to level three alongside a falling curve for the no-evidence and AdStyle baselines. We add a false-positive rate on clean restyled articles, since a detector that simply flags restyling would otherwise look robust.

The streaming check scores the one hundredth article of a developing story. We plot wall-clock time and memory per new article against bundle size from 5 to 100 articles for DeltaNet, the long-context Transformer, and the cross-encoder, and we report the accuracy each achieves at each size. DeltaNet's reported training speedup of up to 30x over sequential recurrence makes chunkwise training feasible. The inference-state advantage has not been measured for this task, so we measure it here.

The expressivity check addresses a result that seems to cut against us. Selective state-space models and similar parallel-scan recurrences sit in the circuit class $TC^0$, informally the class of problems solvable by constant-depth circuits, and so cannot compose long relation chains The Computational Limits of State-Space Models and Mamba via the Lens of Circuit Complexity. A one-hop swap is associative recall with a conflict, which is within reach. The two-hop swap requires chaining two relations. We predict that accuracy on two-hop swaps rises with the number of layers and falls well below the one-hop level at depth one. If instead two-hop accuracy matches one-hop at depth one, the model is probably exploiting a shortcut, and we will audit the benchmark for leakage. Either outcome tells the field where recall-based verification stops working.