Financial misinformation poses severe risks to investors' de- cisions and market stability, and detecting such misinforma- tion in a reference free setting where no external evidence is available remains a critical challenge. This study proposes a rationale guided fine tuning framework for reference free financial misinformation detection, tailored to the ICWSM 2026 Financial Misinformation Detection shared task. The framework first leverages the DeepSeek-V3 to generate two complementary rationales, (text description based and com- monsense based) for each financial news paragraph with- out relying on web search or external knowledge retrieval. These rationales, along with the original news content, are then fed into a BERT-based model,it integrates the powerful reasoning capabilities of LLM with the refined domain spe- cific capabilities of SLM. Experimental results on the official task dataset demonstrate the effectiveness of the proposed ap- proach, achieving an F1-score of 0.7550. This performance outperforms baseline methods that lack rationale guidance, verifying the value of integrating interpretable reasoning into reference free misinformation detection.

PDF