Much of what the public and other scientists hear about Alzheimer's treatments rests on a small number of figures. A typical one is a panel of mouse brain sections stained for a glial marker, alongside a bar chart showing that treated animals have less staining. A press release or a citing paper then turns that panel into a sentence such as "the drug reverses Alzheimer's neuroinflammation." The sentence may be literally supported by the bar chart and still be misleading. A drop in one marker, measured in one brain region of a transgenic mouse, does not show that inflammation is resolved, that the disease is reversed, or that anything happens in people. Current multimodal fact-checkers, meaning systems that judge a claim against an accompanying image, are not built to catch this kind of gap. They look for edited pixels, for captions that do not match the image, or for flipped numbers. They do not ask what the measurement in the image is actually able to establish.
We propose an assay-aware claim verifier. It reads a claim and a scientific figure, finds the relevant panel, and writes both into the same small set of slots: which marker, which brain region, which model system, which comparison, which direction, and whether the difference was significant. It then compares the two slot by slot, using an expert-written table of what each neuroinflammation measurement can and cannot support. The verdict is explained by the specific slot that fails, for example "the claim says 'patients' but the panel is APPswe/PS1dE9 mice," or "GFAP intensity measures astrocyte reactivity, not cognitive rescue." The bet is that these "overreach" errors are where today's vision-language models fail most often. We also expect that better panel localization, the main fix proposed so far, will not repair them.
Core Claim
Errors in verifying biomedical figure claims fall into two kinds. The first is reading errors: the model looks at the wrong panel or misreads a direction or significance marker. The second is inference errors: the model reads the panel correctly but accepts a claim that goes beyond what the assay licenses. Prior work has measured and partly fixed the first kind. MuSciClaims evaluates vision-language models on claims about multi-panel figures from Cell and Nature Physics. It shows that highlighting the gold panel raises GPT-4o's biology F1 from 0.55 to 0.75, and it documents a "yes-man" bias toward "Supported." Its perturbations, however, are reading-level: inverted verbs, flipped comparisons, altered significance.
We claim three things. On Alzheimer's and neuroinflammation figures, a large share of realistic misleading claims are inference-level overreach. Panel localization leaves most of that error in place. An explicit assay schema, applied after slot extraction, removes much of it while producing explanations that domain readers find checkable. The idea fails if a strong reasoning model given gold localization and a careful chain-of-thought prompt matches the schema-based verifier on the overreach categories.
Where This Sits
The proposal takes its explanation style from text-based misinformation work that names the mechanism of deception instead of just outputting a label. Fine-grained propaganda analysis tags spans with techniques such as loaded language. Persuasive-strategy detection for health misinformation uses a strategy taxonomy (including exaggeration and selective omission) as an intermediate step, and reports that giving GPT-4 gold strategy labels raises Macro F1 on Health-MultiFC from 0.904 to 0.933. CoENLI breaks biomedical claim verification into term grounding, evidence comparison and relation prediction. It reports F1 0.86 on NLI4CT with GPT-4o-mini, and shows that its explanations raised medical-student agreement from 0.74 to 0.81. All three are text-only. Our proposal keeps their idea that the explanation should be the deceptive mechanism, but grounds that mechanism in a figure panel.
The multimodal side supplies the structure. NSLM represents three deceptive patterns as latent variables and combines them with first-order logic rules: image manipulation, cross-modal inconsistency, and image repurposing found by reverse image search. The dual-inconsistency network checks content against the Freebase knowledge graph. MMD-Agent runs text, visual and cross-modal checks in sequence. We reuse the idea of a verdict that is a logical function of named patterns, but our patterns are explicit, human-readable slots rather than latent variables. None of these systems has categories for model-system or construct overreach. Their knowledge sources, Freebase and web search, carry no information about what an Iba-1 stain measures. Work on deepfake detectors adds a constraint: pixel-forensics models miss deceptions where the image is authentic. Every overreach case is of this kind, so forensics is not a candidate solution.
The neuroinflammation literature supplies the schema. Its own open problems also constrain it. Different markers answer different questions: Iba-1 reports microglial presence and shape (Guo et al.), CD68 reports phagocytic and lysosomal activity (Cheng et al.), GFAP intensity reports astrogliosis (Wu et al.), C3d and S100A10 co-staining separates A1 from A2 astrocytes (Nie et al.), and CSF sTREM2 is a living-patient proxy with different implications (Franzmeier et al.). The research file also notes that binary A1/A2 or M1/M2 labels oversimplify glial states. The schema must therefore encode what a marker cannot establish, and avoid treating markers as switches. The file also notes that familial transgenic mice do not capture sporadic AD. This is why the model-system slot matters. Studies such as ibrutinib in AD mice or 40 Hz ultrasound in female AD mice give realistic templates for claims that drop species, sex or region on the way to a headline. We do not claim these particular papers were misreported. Finally, AlzKB, a knowledge graph of 118,902 nodes and 1,309,527 edges, lets us check mechanistic links a claim asserts, such as "via Nrf2." Graph triples alone cannot read a figure, which is the gap the slot extractor fills.
Research Program
The first deliverable is a benchmark, tentatively NeuroInflammClaims, built from openly licensed AD and neuroinflammation papers. Its figures contain IHC or immunofluorescence panels and their quantifications for Iba-1, CD68, GFAP, C3d/S100A10, NeuN, MBP, and related cytokine and CSF measures. For each figure, neuroscience annotators write the panel-level evidence record and a faithful claim. They then derive perturbed claims in three families. Reading perturbations follow MuSciClaims: flipped direction, flipped significance, wrong group. Overreach perturbations come in three subtypes. Construct overreach turns a marker change into a disease-level or functional conclusion. Population overreach moves from mice, one sex, or in vitro cells to humans. Mechanism overreach states a causal pathway that the panel only correlates. The third family is out-of-context reuse, in which a real micrograph is paired with a claim about a different study. Labels are Supported, Contradicted, Overreach (with subtype) and Unrelated. Each label carries a gold panel and a gold failing slot. We target roughly 1,500 pairs, comparable to MuSciClaims. Splits are by source paper, so figures never leak across splits. Two annotators label each item, and we report their agreement.
The second deliverable is the verifier, which has three stages. A panel localizer and reader turns the figure into evidence records. A claim parser fills the same slots from the claim text. A comparator applies the assay schema and returns a verdict plus the failing slot. The schema is a short, versioned table written with domain experts. For each marker and readout, it lists the constructs it can support and the ones it cannot. Mechanistic slots are checked against AlzKB paths. The reader and parser are prompted, and optionally lightly fine-tuned, on the same vision-language backbone used by the baselines. This keeps the contribution in the structure, not in the model. The comparator is deterministic, so every verdict can be traced to a slot.
The third deliverable is an explanation study. Biomedical graduate students judge claim-figure pairs with no help, with a free-text chain-of-thought explanation from a strong model, or with our slot-level explanation. We measure accuracy, agreement and time, following the agreement analysis in CoENLI.
Baselines
The naive reference is a classifier that always answers "Supported." Given the documented yes-man bias, this is closer to real model behavior than it sounds. Any system must clearly beat it on macro F1 and on recall of Contradicted and Overreach.
The strongest existing method is a frontier reasoning vision-language model prompted zero-shot with claim and figure. MuSciClaims reports o3 and o4-mini at F1 0.72, the best figure in the file for multimodal scientific claim verification. We run o4-mini (or its closest available successor) and an open model, InternVL3, which MuSciClaims reports at 0.70. Both run on our benchmark and on the MuSciClaims biology subset. Every condition, including ours, gets the same prompt-tuning budget on a shared validation split and the same input resolution. We spend comparable token budgets by allowing the baselines extended reasoning.
A second strong competitor is the same backbone given gold or predicted panel localization. This is the intervention MuSciClaims found most effective, raising GPT-4o biology F1 from 0.55 to 0.75. A third is an agentic pipeline in the style of MMD-Agent, which reaches 61.6% F1 with GPT-4V on MMFakeBench, run with the same backbone and tool budget. A text-only CoENLI baseline sees the claim plus figure caption and legend but no image. It tests whether the figure matters at all.
The key ablation removes the assay schema. The model fills slots and compares them, but judges "does the panel support this conclusion" with its own knowledge. This isolates the schema from the slot structure. To count as beating the strongest competitor, the full system must exceed the gold-localization baseline on Overreach macro F1 by a margin larger than the paired-bootstrap 95% interval. It must also not fall below it on reading perturbations or on the MuSciClaims biology subset. No reported result was found on Lacuna for any model on overreach-style claims about neuroinflammation figures, so all numbers on the new benchmark will be produced in this work.
Why not just…?
Why not just use a stronger reasoning model with gold localization? This is the real threat. Localization was the largest single gain MuSciClaims found, and frontier models know that GFAP is an astrocyte marker. Our counter-hypothesis is that knowing the marker differs from refusing a claim that overreaches from it, and the yes-man bias suggests models accept fluent overreach. The experiment settles this directly. If gold localization plus extended reasoning matches the full system on Overreach F1, the schema is unnecessary. We would then report the benchmark and the error taxonomy as the contribution and drop the verifier claim.
Why not just verify the caption and claim as text, CoENLI-style? Captions often state the marker, species and direction, so much overreach might be caught from text alone. That is useful to know, but captions routinely omit region, sex and significance, and out-of-context reuse cannot be detected without the image. The text-only baseline measures exactly how much of the benchmark is solvable without the figure. We will report per-category results so that any category text solves is flagged honestly.
Why not just retrieve the source paper by reverse image search, as NSLM does? Retrieval handles the out-of-context category well when the original is indexed. The file notes that it fails on novel or private imagery. More importantly, finding the source paper does not decide whether a claim overreaches from it. We include a retrieval-oracle condition, which receives the correct source paper's text, to separate these two effects.
Why not just fine-tune an existing multimodal detector on the new data? MMFakeBench reports that fine-tuning specialized detectors such as HAMMER on new patterns degrades earlier ones, and such detectors give no slot-level reason. We run one fine-tuned open-model baseline on the training split for completeness. We then compare its accuracy and its forgetting on MuSciClaims against our mostly prompted system.
Decisive Checks
The central result is a per-category error decomposition. We expect gold localization to close most of the gap on reading perturbations and little of it on the three overreach subtypes, while the schema closes the overreach gap. If localization closes both, the core claim is wrong.
Faithfulness is tested by intervention. Editing a single slot in the claim, such as changing "mice" to "patients," should flip the verdict to Overreach for the stated reason in at least as many cases as it does for the baselines' free-text rationales. We measure the agreement between the predicted failing slot and the gold slot, and the agreement between the predicted panel and the gold panel.
A transfer check runs the system unchanged on the MuSciClaims biology subset. It tests whether the slot structure helps beyond neuroinflammation or overfits to it. The human study then asks the practical question: do slot-level explanations raise reviewer accuracy and agreement on overreach items beyond what free-text chain-of-thought provides? CoENLI's 0.74 to 0.81 agreement gain is the text-only reference point.
References
- MuSciClaims: Multimodal Scientific Claim Verification
- Fine-Grained Analysis of Propaganda in News Articles
- Using Persuasive Writing Strategies to Explain and Detect Health Misinformation
- Explainable Biomedical Claim Verification with Large Language Models
- Unveiling Implicit Deceptive Patterns in Multi-Modal Fake News via Neuro-Symbolic Reasoning
- Inconsistency Matters: A Knowledge-guided Dual-inconsistency Network for Multi-modal Rumor Detection
- MMFakeBench: A Mixed-Source Multimodal Misinformation Detection Benchmark for LVLMs
- Fact or Fake? Assessing the Role of Deepfake Detectors in Multimodal Misinformation Detection
- Optogenetic Activation of the Excitatory Neurons Expressing CaMKIIαin the Ventral Tegmental Area Upregulates the Locomotor Activity of Free Behaving Rats
- LRRK2 Deficiency Aggravates Sleep Deprivation-Induced Cognitive Loss by Perturbing Synaptic Pruning in Mice
- A modified mouse model of perioperative neurocognitive disorders exacerbated by sleep fragmentation
- Increasing brain N‐acetylneuraminic acid alleviates hydrocephalus‐induced neurological deficits
- Higher CSF sTREM2 attenuates ApoE4-related risk for cognitive decline and neurodegeneration
- Chronic oral administration of ibrutinib prevents long-term memory deficits and reduces AD pathology and neuroinflammatory responses in a mouse model of AD
- 40 Hz transcranial ultrasound stimulation improves cognition and neuroinflammation in female Alzheimer's disease mice
- The Alzheimer’s Knowledge Base: A Knowledge Graph for Alzheimer Disease Research