We present an overview of the Reference-Free Financial Misinformation Detection shared task at the 2nd Workshop on Misinformation Detection in the Era of LLMs (MisD). Built on RFC-Bench, the task requires systems to classify a single financial news paragraph as factual or minimally manipulated, without access to any external evidence or source document. This reference-free setting islocates the difficulty of reasoning about internal semantic consistency and financial plausibility. The task attracted multiple participating teams whose approaches span three paradigms: parameter-efficient LLM adaptation, discriminative learning with controlled augmentation, and reasoning-augumented multi-agent frameworks. On the private test set, the top system reaches an F1 of 0.9629, while a 2-shot GPT-4.1 baseline remains close to chance at 0.5686. The results show clear progress on pattern-level detection, but also indicate that reference-free reasoning from a single paragragh remains an open problem.

PDF