Researchers have developed a new method to mitigate hallucinations in large language models. Hallucinations occur when the model generates responses that sound plausible but are factually incorrect. The new approach, called Responsible Contrastive Soft Prompting (RCSP), uses soft prompts to balance three goals: suppressing hallucinated content, encouraging the model to abstain from answering if it's unsure, and preserving factual recall. RCSP achieves these goals by incorporating contrastive loss, curriculum learning, and KL regularization into its training mechanism. The method was evaluated on five generative question-answering datasets using an LLM-as-a-Judge framework, with results showing that RCSP effectively balances factual recall with hallucination suppression and abstention, outperforming standard reasoning and instruction-based prompting baselines. --- Why it matters: This matters to engineers working on large language models because it provides a new approach to improving model reliability in high-stakes domains. By mitigating hallucinations, RCSP can help reduce errors that undermine trust and introduce real-world risk. Source: https://arxiv.org/abs/2606.00919