Technical Summary: RW-Post and AgentFact

Problem Statement

The proliferation of social media has accelerated the spread of persuasive multimodal misinformation, where manipulated or repurposed images are used to reinforce misleading text, posing significant risks to public health and social behavior. While recent generative AI advances have lowered the cost of producing deceptive content, existing automated verification systems face critical limitations. Current approaches often focus on narrow subtasks (e.g., out-of-context image detection or deepfake detection) or rely on Large Language Models (LLMs) and Large Vision-Language Models (LVLMs) with shallow evidence usage. A primary challenge is faithfulness: models frequently generate fluent rationales that hallucinate reasoning or misuse evidence, lacking grounding in verifiable sources. Furthermore, the field lacks benchmarks that jointly provide real-world multimodal claims, auditable reasoning traces, and explicitly linked evidence derived from human fact-checking articles.

Methodology

1. RW-Post: A Post-Aligned Multimodal Fact-Checking Benchmark

To address the gap in existing datasets, the authors introduce RW-Post, a benchmark designed for explainable, evidence-grounded multimodal fact-checking.

• Data Construction: The dataset is constructed from 13,000 Snopes fact-checking articles (2017–2024). The authors employ an LLM-assisted extraction-and-auditing pipeline to reformat HTML articles into a "text-with-tagged-URL" representation. This preserves the contextual position of links, allowing LLMs to reliably extract original social media posts, images, reasoning traces, and evidence.
• Annotation & Quality Control:
  • Relabeling: Original Snopes labels are consolidated into five high-level categories (True, False, Unproven, Mixture, Outdated) to ensure consistency.
  • Filtering: Automated and human verification steps remove articles with insufficient text, low-resolution images, or "label leakage" (e.g., images containing watermarks like "fake" or "false").
  • Evidence Alignment: The dataset explicitly links reasoning key points to specific evidence items (text, images, videos) with stable IDs and URLs.
• Evaluation Regimes: RW-Post supports three complementary evaluation modes:
  1. Closed-book: Models observe only the post text and image.
  2. Evidence-bounded: Models receive the post plus a fixed set of oracle evidence items from the dataset.
  3. Open-web: Models retrieve evidence via web search under a standardized, leakage-controlled protocol (filtering out known fact-checking sites and post-dating sources).

2. AgentFact: A Reference Verification Pipeline

The authors propose AgentFact, a modular, agent-based framework to serve as a reference baseline for open-web verification. It decomposes the verification task into five specialized agents operating in an iterative retrieve-reason workflow:

1. Strategy Planning Agent (Agent-SP): Generates a verification plan, including reasoning steps, a validation list, and search intents.
2. Text Evidence Retrieval Agent (Agent-TR): Retrieves web evidence, filters results, and assesses source reliability (categorizing sources as reliable, unreliable, satire, or unsure).
3. Image Retrieval and Analysis Agent (Agent-IR): Performs reverse image search to detect visual inconsistencies, including near-duplicates, same-event images, visual tampering, and miscaptioning.
4. Reasoning Agent (Agent-R): Integrates the claim, post context, and retrieved evidence to perform step-wise reasoning, producing a veracity label and confidence score.
5. Explanation Generation Agent (Agent-EG): Synthesizes the reasoning and evidence into a structured explanation with explicit citations.

Key Contributions

1. RW-Post Benchmark: The introduction of a real-world multimodal fact-checking benchmark that aligns claims with original posts and provides auditable reasoning traces and source-grounded evidence. It is the first to offer explicit evidence-reasoning alignment for auditing.
2. Extraction Pipeline: A novel LLM-assisted pipeline for extracting and auditing reasoning and evidence from human-written fact-checking articles, ensuring high-quality, structured annotations.
3. AgentFact Framework & Baselines: The provision of AgentFact as a reference verification agent and the benchmarking of strong open-source and closed-source LVLMs under unified closed-book, evidence-bounded, and open-web protocols.

Experimental Results

The authors evaluated AgentFact and various LVLM baselines (LLaVA-1.5, Qwen2-VL-Chat, GPT-4o-mini) across multiple datasets, including RW-Post, Mocheg, CLAIMREVIEW2024+, and NewsCLIPpings.

• Open-Web Performance: AgentFact achieved the best performance among fact-checking systems on RW-Post and CLAIMREVIEW2024+ in terms of Macro-F1, Weighted-F1, and Accuracy, outperforming retrieval-augmented baselines like LEMMA and DEFAME.
• Impact of Evidence: In controlled evaluations on RW-Post, providing evidence (Evidence-bounded setting) substantially improved performance for all models. For instance, GPT-4o-mini's Weighted-F1 increased from 25.4 (closed-book) to 85.1 (evidence-bounded).
• Visual Grounding Challenges: Adding images to the input (T+I or T+I+E) yielded limited or negative gains compared to text-only or text-plus-evidence inputs. This suggests current LVLMs struggle to effectively integrate visual signals with textual evidence, often failing to utilize the image for verification when evidence is available.
• Faithfulness and Hallucination: Human evaluation revealed that while ground-truth rationales are highly faithful, model outputs suffer from reasoning and evidence usage hallucinations. AgentFact significantly outperformed LEMMA in faithfulness scores, yet still lagged behind human annotations.
• Format Compliance: Open-source LVLMs showed significantly lower Format Success Rates (FSR) compared to GPT-4o-mini, indicating that generating structured, evidence-grounded outputs remains a non-trivial challenge for open models.

Significance and Claims

The paper argues that RW-Post reveals substantial "headroom" for multimodal verification research. The experiments demonstrate that while current models struggle with faithful evidence grounding and visually informed reasoning, the evidence-bounded evaluation significantly improves both accuracy and faithfulness. The authors claim that RW-Post provides a necessary testbed for systematically diagnosing visual grounding and evidence utilization, moving beyond shallow matching to rigorous, auditable fact-checking. The work highlights that reliable evidence is the dominant factor in verification performance, while robust multimodal refutation and contextual grounding remain open challenges. The dataset and code are released to facilitate future research in evidence-aware multimodal fact-checking.