FactGuard: Agentic Video Misinformation Detection via Reinforcement Learning

A team of researchers led by Zehao Li and including nine co-authors introduced FactGuard, an innovative agentic framework for detecting video misinformation through reinforcement learning, in a paper published on the arXiv preprint server on February 26, 2026, aiming to overcome limitations in current multimodal large language models that struggle with sparse or fragmented evidence. The research addresses a critical challenge in digital media verification, as existing multimodal large language models (MLLMs) often rely on fixed-depth inference and place excessive trust in internally generated assumptions, particularly when critical evidence is limited or requires external verification. FactGuard represents a significant advancement by formulating verification as an iterative reasoning process built upon MLLMs, which explicitly assesses task ambiguity and selectively invokes external tools to acquire critical evidence, enabling progressive refinement of reasoning trajectories. The researchers implemented a sophisticated two-stage training strategy that combines domain-specific agentic supervised fine-tuning with decision-aware reinforcement learning to optimize tool usage and calibrate risk-sensitive decision making. This approach allows the system to dynamically determine when additional verification is needed rather than making conclusions based solely on initial assumptions. Extensive experiments conducted on three benchmark datasets—FakeSV, FakeTT, and FakeVV—demonstrate FactGuard's state-of-the-art performance and validate its excellent robustness and generalization capacity across different types of video misinformation. The framework's ability to adaptively seek additional evidence when needed represents a paradigm shift from static verification approaches to more dynamic, evidence-based reasoning processes that mirror human fact-checking methodologies. This research has significant implications for the growing field of AI-powered misinformation detection, particularly as deepfakes and manipulated videos become increasingly sophisticated and prevalent in social media and news dissemination. The FactGuard framework's agentic approach—where the system can actively seek additional information rather than passively processing input—could serve as a foundation for next-generation content verification systems. The open publication of this research on arXiv allows the broader AI research community to build upon these findings, potentially accelerating the development of more reliable tools for combating digital misinformation in an era where distinguishing authentic from manipulated content has become increasingly challenging.