Large Language Models (LLMs) often generate hallucinations, producing outputs that are contextually inaccurate or factually incorrect. We introduce HICD, a novel method designed to induce hallucinations for contrastive decoding to mitigate hallucinations. Unlike existing contrastive decoding methods, HICD selects attention heads crucial to the model’s prediction as inducing heads, then induces hallucinations by dispersing attention of these inducing heads and compares the hallucinated outputs with the original outputs to obtain the final result. Our approach significantly improves performance on tasks requiring contextual faithfulness, such as context completion, reading comprehension, and question answering. It also improves factuality in tasks requiring accurate knowledge recall. We demonstrate that our inducing heads selection and attention dispersion method leads to more “contrast-effective” hallucinations for contrastive decoding, outperforming other hallucination-inducing methods. Our findings provide a promising strategy for reducing hallucinations by inducing hallucinations in a controlled manner, enhancing the performance of LLMs in a wide range of tasks.

Adverse drug events (ADEs) pose major public health risks, with traditional reporting systems often failing to capture them. Our proposed pipeline, called Deep-LLMADEminer, used natural language processing approaches to tackle this issue for #SMM4H 2024 shared task 1. Using annotated tweets, we built a three part pipeline: RoBERTa for classification, GPT-4-turbo for span extraction, and BioBERT for normalization. Our models achieved F1-scores of 0.838, 0.306, and 0.354, respectively, offering a novel system for Task 1 and similar pharmacovigilance tasks.