How does DEFAME perform multimodal fact-checking with dynamic evidence?

Why checking images needs image evidence

The paper starts from a practical gap: automated fact-checking systems are mostly text-only, and only a few handle visual claims. Even those rarely handle multimodal claims and multimodal evidence at once, and most multimodal systems cannot retrieve the evidence needed to verify a claim. The authors cite that roughly 80% of claims checked by professional fact-checkers are multimodal, which makes this gap consequential rather than cosmetic.

DEFAME's answer is to treat the claim image as part of the context instead of converting it into a text description. The paper contrasts this with RAGAR, a close relative that processes textual and visual claims but retrieves only textual evidence and turns the claim image into a description, discarding visual information. DEFAME instead retrieves and evaluates both textual and visual evidence, and returns a human-readable report rather than a bare label.

Six stages with dynamic tool choice

DEFAME combines a multimodal LLM, a suite of multimodal tools, and a structured fact-checking report, operating as a dynamic multi-step retrieval-augmented system inspired by established fact-checking workflows. The process is decomposed into six manageable stages. Each call to the MLLM includes the current state of the report as context plus a task-specific description, which the authors describe as a form of context awareness that guides the model to pertinent information and supports multi-hop reasoning and evidence retrieval.

The tools are not decorative. In the paper's example reports, the system calls reverse image search, web search, image search, and geolocation. For a claim about a "basket man" in 1960s Istanbul, the geolocation tool returns a country probability list with Turkey highest at 0.26, and image search returns contextual discussion of the practice. The dynamic part is that the system selects tools and search depth per claim rather than following a fixed retrieval recipe.

What the three benchmarks measure

The evaluation spans three benchmarks with different demands. AVerITeC is used because the authors say it mitigates weaknesses of earlier benchmarks, including reliance on artificial claims and evidence insufficiency. MOCHEG builds on real-world claims, requires multimodal evidence retrieval, and adds justification generation. VERITE is used because it reduces unimodal bias and incorporates real-world samples relative to earlier out-of-context image benchmarks.

These are not interchangeable tests. MOCHEG stresses retrieval plus explanation, VERITE stresses out-of-context image detection with real-world samples, and AVerITeC stresses more realistic claim and evidence conditions. The paper's claim of state of the art covers uni- and multimodal fact-checking across these benchmarks, so the result should be read as breadth across task types rather than a single score on one dataset.

A leakage-free test of recency

The paper introduces ClaimReview2024+, a multimodal benchmark of claims posted after the GPT-4o knowledge cutoff, built to avoid data leakage. Claims come from the Google Fact Check Tools API and ClaimReview, with label assignment in two stages: trivial cases mapped automatically by an LLM-based script, and ambiguous ones manually labeled by a PhD-level AFC researcher, then validated by a second annotator against the original fact-checking articles.

The final label distribution is 129 Refuted, 89 Supported, 61 Misleading, and 21 NEI. The authors manually rewrote claims containing explicit fact-checking language to prevent leakage of veracity judgments, and manually curated images because the API often returns teaser images with overlays or composites. They also note a caveat: all claims are posted after the cutoff, but some may describe earlier events re-evaluated by fact-checking organizations.

Where DEFAME beats GPT-4o most

On ClaimReview2024+, the paper reports that DEFAME drastically outperforms the GPT-4o baselines. The authors interpret this as evidence of temporal generalizability and potential for real-time fact-checking. The mechanism they point to is that DEFAME uses internal knowledge mostly for commonsense reasoning and retrieves evidence dynamically through external tools, whereas systems relying on parametric knowledge have static knowledge that fails on recent claims.

That comparison is the paper's sharpest practical distinction. A model answering from memory cannot see a claim posted after its cutoff, while a retrieval-based pipeline can search for it. The paper also argues that parametric-only predictions lack links to verifiable sources and raise hallucination risk. The ClaimReview2024+ result is therefore a test of the retrieval design, not just of raw model capability.

Reports explain, but retrieval can fail

DEFAME's output is a structured report with a verdict, justification, evidence, and the actions taken. The paper's appendix shows both successes and failures. In one wrong prediction, the system judged a claim about an Air Force Officers' Spouses' Club photo as supported after finding evidence about the event, but the authors attribute the error to a premature judgment. In another, repeated searches returned "No new evidence found," and the system labeled a fishnet-on-power-lines claim as out of context; the authors attribute that error to failed retrieval of evidence.

These examples matter because they locate the failure mode in the pipeline rather than in the verdict wording. The limitations section adds related risks: reliance on search engines can incorporate unreliable or biased content, web scraping is prone to instability from restricted access and large documents, and open-source models such as LLaVA-1V show formatting sensitivity where minor prompt variations affect response quality. Hallucination remains an inherent LLM risk, though the authors say qualitative analysis and human evaluation did not indicate substantial amounts.

What the result does not yet support

The paper is explicit that the results do not sufficiently warrant real-world application yet. It frames DEFAME as a step toward explainable, verifiable AI and as an augmentation tool for human fact-checkers rather than a black-box authority. The authors also warn that automated fact-checking can shape narratives by prioritizing certain sources or amplifying institutional biases in retrieval mechanisms or parametric knowledge, and that governments or platforms could misuse such systems for overreach in content moderation.

The backbone-agnostic design is presented as a partial answer: DEFAME would automatically benefit from efforts to de-bias LLMs. But the authors concede that reliability inherently depends on backbone quality, which is difficult to rigorously analyze. For readers weighing the result, the relevant boundary is that the strong ClaimReview2024+ showing tests recency and retrieval, while source credibility, scraping stability, and human oversight remain open conditions for deployment.

Sources used in this answer