مشخصات پژوهش

Automatic fake news detection in Persian is challenging due to limited annotated resources and the subtle boundary between misinformation and legitimate content. It remains unclear whether zero shot prompting of large language models alone suffices or whether task specific adaptation is necessary. This paper empirically compares classical lexical baselines, zero shot instruction tuned LLMs, fine tuned ParsBERT, and LoRA tuned LLMs (Qwen3 8B and Dorna Llama3 8B) on a Persian fake news benchmark. Results show that zero shot LLMs substantially underperform TF IDF baselines, exhibiting poor calibration and precision recall trade offs. However, supervised adaptation dramatically changes this conclusion: LoRA tuned Qwen3 8B achieves a fake class F1 of 0.745, ParsBERT reaches 0.744, and a validation trained stacker over the fine tuned models yields the best overall performance with a fake class F1 of 0.761 and accuracy of 0.799. These findings demonstrate that zero shot prompting is insufficient for this task, but parameter efficient fine tuning makes local LLMs highly competitive for Persian fake news detection.