Reference-free financial misinformation detection asks whether a standalone paragraph is authentic or semantically perturbed without retrieving external evidence. The shared task is challenging because misleading examples can re- main fluent while differing from authentic financial news by only a small numeric, directional, or attributional edit. We describe a two-stage system that combines LLM-guided data augmentation with a multi-task encoder-based classifier. Starting from real seed paragraphs, an LLM extracts com- mitments, generates label-changing perturbations and label- preserving controls, and filters candidates with overlap, edit- ratio, and leakage checks. The ROBERTA-large encoder is trained with binary classification, perturbation-family pre- diction, novelty detection, cue-span supervision, contrastive learning, and semantic-group ranking. An optional reranker consults the LLM only on uncertain cases. On the develop- ment split, the full system achieves 95.55% accuracy and 95.59% macro-F1 on the blind data set. The largest gains come from benign controls and groupwise ranking, while at- tributional and compositional edits remain challenging.

PDF