Technical Summary: RW-Post – Auditable Evidence-Grounded Multimodal Fact-Checking in the Wild

1. Problem Statement

The proliferation of social media has accelerated the spread of persuasive multimodal misinformation, where images and videos are manipulated or repurposed to reinforce misleading text, posing significant risks to public health and social behavior. While generative AI further lowers the cost of producing realistic deceptive content, existing automated verification systems face critical limitations:

• Shallow Evidence Usage: Current approaches often rely on shallow matching or lack faithful grounding in verifiable sources.
• Lack of Auditable Benchmarks: There is a scarcity of benchmarks that jointly provide real-world multimodal claims, auditable reasoning traces, and explicitly linked evidence.
• Context Loss: Many datasets fail to preserve the original social media post context, which is critical for accurate verification.
• Evaluation Gaps: Existing methods struggle with "faithfulness," where models generate fluent but hallucinated rationales unsupported by verifiable sources, particularly in open-web settings.

2. Methodology

A. The RW-Post Benchmark

The authors introduce RW-Post, a post-aligned text–image benchmark designed for real-world multimodal fact-checking with auditable annotations.

• Data Construction: The dataset is derived from 13,000 Snopes fact-checking articles (2017–2024).
• Extraction Pipeline: An LLM-assisted extraction-and-auditing pipeline converts HTML articles into a "text-with-tagged-URL" representation. This preserves the contextual position of links, allowing the model to distinguish between the original post, images, reasoning, and external evidence.
• Annotation Structure: Each instance links:
  1. The original social media post (text and image).
  2. Structured reasoning traces (logic and key points).
  3. Explicitly linked evidence items (with stable IDs and source URLs) derived from human-written fact-check articles.
• Quality Control: The pipeline includes multimodal relevance filtering (removing claims where images are irrelevant), watermark leakage filtering (removing images with "fake" or "false" labels), and human validation to ensure reasoning and evidence alignment.
• Evaluation Regimes: RW-Post supports three complementary evaluation settings:
  1. Closed-book: Models observe only the post and image.
  2. Evidence-bounded: Models receive the post plus a fixed set of oracle evidence items.
  3. Open-web: Models retrieve evidence via web search under a standardized, leakage-controlled protocol (filtering out known fact-checking sites and post-dated sources).

B. Reference Verification Pipeline: AgentFact

To establish baselines, the authors propose AgentFact, a modular, agent-based verification framework that integrates strategy planning, evidence retrieval, visual analysis, reasoning, and explanation generation.

• Agent-SP (Strategy Planning): Generates a verification plan, including validation lists and search intents.
• Agent-TR (Text Evidence Retrieval): Retrieves and filters textual evidence, assessing source reliability (reliable, unreliable, satire, unsure).
• Agent-IR (Image Retrieval & Analysis): Performs reverse image search to detect visual tampering, miscaptioning, and deepfakes.
• Agent-R (Reasoning): Executes the verification plan step-by-step, grounding reasoning in retrieved evidence and assessing confidence.
• Agent-EG (Explanation Generation): Produces a structured explanation with explicit evidence-ID citations.
• Workflow: The system operates on an iterative retrieve–reason loop, alternating between evidence acquisition and evidence-grounded reasoning until a veracity prediction is reached.

3. Key Contributions

1. RW-Post Benchmark: A new benchmark for explainable multimodal fact-checking featuring post-aligned claims, reasoning traces, and source-grounded evidence, addressing the gap in existing datasets that lack original context or auditable evidence links.
2. Extraction Pipeline: An LLM-assisted pipeline for extracting and auditing reasoning and evidence from human-written fact-checking articles, ensuring structured and verifiable annotations.
3. AgentFact Framework & Baselines: A reference verification agent and a comprehensive benchmarking of strong open-source Large Vision-Language Models (LVLMs) under controlled closed-book, evidence-bounded, and open-web protocols.

4. Experimental Results

A. Open-Web End-to-End Performance

• Fact-Checking (3-class): AgentFact achieved the best performance among fact-checking systems on RW-Post, ClaimReview2024+, and Mocheg in terms of Macro-F1, Weighted-F1, and Accuracy. It demonstrated superior handling of class imbalance compared to retrieval-augmented baselines like LEMMA and DEFAME.
• OOC Detection (Binary): On the NewsCLIPpings dataset, specialized OOC detectors (CCN, Sniffer) outperformed general-purpose fact-checking systems, highlighting the distinct nature of out-of-context detection versus open-web fact-checking.

B. Controlled Benchmarking (LVLMs)

• Evidence Impact: Providing evidence (T+E) substantially improved performance for all models. For instance, GPT-4o-mini's Weighted-F1 jumped from 25.4 (closed-book) to 85.1 (evidence-bounded). Open-source LVLMs (LLaVA-1.5, Qwen2-VL-Chat) also saw significant gains but remained far below closed-source baselines.
• Visual Grounding Challenges: Adding images to the input (T+I or T+I+E) yielded limited or even negative gains in evidence-bounded settings. This suggests current LVLMs struggle to effectively integrate visual signals with textual evidence, often failing to ground visual analysis in the provided context.
• Refutation and Uncertainty: Models performed better at confirming supported claims (True) than rejecting misinformation (False) or handling "Unproven" cases, indicating difficulties in calibrated abstention.
• Format Compliance: Open-source LVLMs showed significantly lower Format Success Rates (50–86%) compared to GPT-4o-mini (100%), indicating that generating structured, evidence-grounded outputs remains a non-trivial challenge.

C. Explainability and Human Evaluation

Human evaluation of 40 sampled instances revealed:

• Ground Truth: Human-annotated rationales achieved near-perfect faithfulness scores (0.94–1.00).
• Model Performance: AgentFact significantly outperformed LEMMA in reasoning faithfulness, evidence usage, and label justification (p < 10⁻⁵).
• Failure Modes: Case studies highlighted "weak contextual grounding," where models retrieved relevant-sounding text but failed to capture the true event-level context of images (e.g., misidentifying the specific event or participants).

D. Ablation Study

Removing the Text Evidence Retrieval Agent (Agent-TR) caused the largest performance drop, confirming it as the most critical component. Removing the Strategy Planning Agent or Image Analysis Agent also resulted in measurable declines, validating the complementary roles of the modular agents.

5. Significance and Claims

The paper claims that RW-Post provides a necessary testbed for advancing research on evidence-aware multimodal fact-checking. The experiments demonstrate substantial "headroom" for current models:

• Evidence is Dominant: Reliable evidence grounding is the primary driver of performance; without it, even strong models struggle.
• Visual Grounding is Hard: Despite the multimodal nature of the task, current LVLMs fail to effectively utilize visual signals when combined with textual evidence, often relying on surface-level text matching.
• Faithfulness Matters: The benchmark exposes the prevalence of hallucinated reasoning and misused evidence in current systems, emphasizing the need for evaluation protocols that audit the link between reasoning and verifiable sources.

The authors conclude that while evidence-bounded evaluation improves accuracy and faithfulness, robust multimodal verification in the wild remains an open challenge, particularly regarding visual grounding and the handling of uncertainty.